Spatial Learning and Overfitting in Visual Recognition and Route

Planning Tasks

Margarita Zaleshina

1 a

and Alexander Zaleshin

2 b

Moscow Institute of Physics and Technology, Moscow, Russia

Institute of Higher Nervous Activity and Neurophysiology, Moscow, Russia

Keywords: Spatial Cognition, Navigation, Machine Learning, Overfitting, Retraining.

Abstract: Spatial data recognition, navigation based on localized visual clues and ability to identify significant elements

in the environment and build routes is formed as a result of general spatial learning and then adjusted to a

specific location. Modern artificial intelligence (AI) — from visual processing applications to autonomous

vehicles—also includes this capability. However, excessive learning can lead to overfitting, which

significantly reduces the efficiency of spatial actions. In this work we describe typical algorithms for

navigation, spatial learning in pigeon flights, and remote sensing recognition in neural networks. We consider

learning algorithms based on significant topological elements, and suggest possible methods to expand

learning opportunities and reduce the impact of erroneous settings. Our calculation results show how

overfitting affects navigation behaviour and visual recognition. Result of this work provides direction for the

future development of new algorithms that optimize the efficiency of spatial learning.

1 INTRODUCTION

Spatial behavior is determined by the internal

settings, goals, and expectations of subjects and by

the external world as well as by ways of obtaining

additional information about the external

environment. The ability to solve spatial problems is

used directly in everyday life and affects global

processes of settlement and migration.

Animals solve their spatial tasks reflexively,

without a detailed study of topological relationships;

they directly relate their observations, actions, and

results to the physical capabilities of their bodies and

the available environment. Humans have an

opportunity to apply both their natural skills and

digital technologies to solve localized problems in

spatial structures (Freksa et al. 2017). An application

of artificial intelligence (AI) complements the

possibilities of spatial perception and navigation of

humans and animals and builds a new level for

discoveries and achievements in geography (Galvani,

Zaleshina, and Zaleshin 2021).

__________________________

https://orcid.org/0000-0001-5273-6579

https://orcid.org/0000-0001-9356-9615

Detailed skills for terrain orientation are acquired

through spatial learning. There are different ways to

evaluate the effectiveness of spatial actions. In

practice, it can be summarized into two main

indicators - how many wanted objects of spatial

search are found and reached, and how much time and

material resources are spent. Additionally, for AI

often calculates a percentage of correct finds in

relation to all finds, and a percentage of found objects

in relation to all objects. This work compares spatial

perception and navigation behavior for animals

(using the example of pigeons), and for artificial

neural networks (using the example of recognizing

basic urban landmarks). Particular attention is paid to

the issues of overfitting and underfitting. It is

emphasized that learning increases speed and

minimizes costs of wayfinding, but at the same time,

overfitting and lack of updates to the applied patterns

lead to systematic repeated errors, leading to a large

number of false results during actions.

The materials of this work can be useful in various

topics related to the formation of spatial cognition and

navigational behavior algorithms.

576

Zaleshina, M. and Zaleshin, A.

Spatial Learning and Overﬁtting in Visual Recognition and Route Planning Tasks.

DOI: 10.5220/0013001100003837

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 16th International Joint Conference on Computational Intelligence (IJCCI 2024), pages 576-583

ISBN: 978-989-758-721-4; ISSN: 2184-3236

2 BACKGROUND

2.1 Spatial Learning Algorithms

The ability to perceive and recognize spatial data and

use this data during movement is innate in both

animals and humans. Now this need is being built into

various AI applications, from programs for remote

sensing recognition to autonomous vehicles. For both

living organisms and AI, learning can be two types:

independent learning and learning with a teacher.

Spatial learning serves to optimize expenditures

of time, effort, and resources. At the same time, the

costs of the path are non-linearly related to the

quantity and quality of the result obtained. In

addition, material resources can be replenished along

the way. Typical learning flowchart is shown in

Figure 1.

Figure 1: Learning stages and typical underfitting and

overfitting problems.

Algorithms, once formed during training, can be

repeated for a long time, but this time is limited. First,

the environment can be modified, e.g., due to climate

change. Second, a repeated action can be a reason of

outside changes, e.g., if the sheep eat and trample all

the grass in the pasture. Third, learning errors can

have a cumulative effect due to filters or due to

repetition.

Another problem that occurs during learning is

overfitting. In the case of a stable environment,

trained algorithms usually work. Overfitting occurs

when a trained algorithm is able to process a limited

set of data in too much detail, it processes well in

fixed sets of samples, but is unable to generalize the

processing to new situations. In fact, novelty is

contraindicated for it. This needs to be eliminated by

retraining on new samples, unlocking the rigidity of

the predefined classification (Bashir et al. 2020).

Additionally, in reinforcement learning, the novelty

of an event can strengthen a motivated response to it

(Siddique et al. 2017), which partially facilitates

efforts to get rid of existing incorrect algorithms.

Overfitting and underfitting can greatly affect

learning outcomes: data can be misinterpreted or

misclassified, relevant data can be filtered out.

Underfitting often results from misinterpretation of

noise or confusion in scales. Some of the noise can be

removed at the preprocessing stage using the decision

tree method (Alharbi 2024) or by segmenting spatial

data according to the scale.

2.2 Typical Spatial Tasks

▪ Bird Flights: Homing and Foraging

In flight, birds focus their attention on the main

elements of the environment: long roads and rivers,

or well-observed objects. Blaser et al. (Blaser et al.

2013) describe how pigeons' mental map helps them

find a route to home or to feeding site based on their

current location. Detailed training in fixed route flight

can be done with a teacher, while flying in a flock, or

while following a leader. Flight consistency in a flock

of pigeons often depends on the flying experience

and/or age of the individual birds and the ability of

the flock leader to set the direction of flight (Santos et

al. 2014).

The difference in the length of the flight paths of

a bird making its first flight and a bird that has

undergone training can differ several times over the

distance A trained pigeon does not need to make a

choice every time where to fly, and its route is

optimally shortened.

Both in the case of the first flight and in the case

of failure to find the original target, the pigeons begin

to survey around the starting point or predicted POI,

over distances comparable to their usual flights.

Schaffner et al. (Schiffner et al. 2018) studied the

difference between surveying trajectories of pigeons

at the moment of departure from the site and

subsequent directed flight to the target. The authors

suggested that complex perception of external

information in pigeons slows down their flight speed,

but at the same time allows the birds to more steadily

and efficiently adhere to the target direction.

Spatial Learning and Overﬁtting in Visual Recognition and Route Planning Tasks

577

▪ AI: Machine Learning and Unlearning

In remote sensing recognition tasks, including for

navigation, the priority is to identify key objects -

roads, buildings, vegetation. Depending on data

sources, quality, and specificity for a particular area,

recognition results of trained neural networks have

different quality. U-Net (Benedetti, Femminella, and

Reali 2022) or DeepLabV3 (Wang et al. 2022) are

often used to recognize remote sensing images. Such

neural networks often have difficulty eliminating

noise interference, such as shadows from trees, when

recognizing buildings or roads. Well-chosen

segmentation labeling algorithm (Lee et al. 2022)

helps to optimize neural networks.

Machine learning models can adapt to variability

in the data they process, although this often requires

pre-training or retraining. In primary sequential

training, the observed data sets are used to fit the

model, assuming that the predicted features are

constant over time. Retraining requires the ability to

collect new data and compare the recognition results

of incoming data with predicted ones (Dietterich

2002). Active learning algorithm improves the model

quality by checking of data labeling and label

dispersion (Bengar, Raducanu, and van de Weijer

2021). Adversarial learning methods are limited in

data generalization and give unreliable results after

overfitting (Zhao, Alwidian, and Mahmoud 2022).

Machine unlearning is used to partially eliminate

incorrect settings, while preserving the neural

network model’s ability to recognize the necessary

data. When unlearning significant indicators that need

to be forgotten and those that need to be remembered

are determined (Foster, Schoepf, and Brintrup 2024).

Hopkins et al. (Hopkins et al. 2024) propose a

model-independent solution based on the ability to

generalize properties across sets of different classes.

Kim et al. (Kim, Kim, and Bengio 2021) associate

each branch of models with a visual concept and

further manage the resulting set using the attention

module. Processing first calculates the content, which

is then returned to a pixel space containing the subject

area and style.

2.3 Route Planning

▪ Aggregation of Route Planning Information.

The rapid development of digital technologies

contributes to a significant increase in the volume of

collection and processing of spatial and temporal

route data. Depending on the structure, route data can

be divided into explicit entities directly related to

observation and implicit additions with weak

spatiotemporal continuity (Kong et al. 2018).

The availability of route points with known

attributes serves as the basis for creating a route

through such natural objects that have been located

close to each other for a long time and usually have

similar or dependent components that determine their

structure and content. To a lesser extent, this applies

to artificial objects. Tobler’s first law (Tobler 1970)

assumes the dependence of some attributes of objects

that are close to each other.

Unlike static orientation elements, dynamic

elements are not constant in their properties over

time. Natural objects can change their visual

properties depending on the time of day and season.

Artificial objects can change their other attributes

without changing visually over time: public facilities

(museums, cafes, shops, etc.) have opening hours;

public transport runs on a schedule, possibly, with

long breaks.

In an unfamiliar environment, a person searches

for previously encountered objects and signs to

recognize other ones. The uncertainty generated by a

little-known situation results in an attempt to orientate

and search for fragments of previously encountered

elements (Tversky and Kahneman 1974). The variety

of identification and interpolation options leads to the

creation of both copies and complementary

extensions of existing fragments. Like puzzles, such

identified fragments do not always form a

recognizable whole. Fragments that are not combined

into blocks collectively make up a potentially usable

pool. The search allows for identifying suitable

fragments and supplements thereto. Missing points

can be added based on the available parts of other

objects, when points with known attributes transfer

their properties onto fragments or whole areas, as is

the case in kriging.

Points with known attributes can also serve as

reference points. Reference points have stable

locations, but their locations can change over time. A

set of reference points forms a system of

spatiotemporal relations that can be used for

orientation along a route. Reference points with

known attributes make it possible to create

generalized coordinate systems based on their spatial

positions or on non-numeric indicators. The complex

nature of reference points allows them to be used as a

tool for operations with objects and attributes, and as

a framework for spatial positioning. The relative

positions of points and objects form the structural

code of the track points. With small changes, the

structural code may remain the same, with significant

changes in the data set of the environment; a new

structural code is formed from some stable or

repetitive components/elements of the environment.

NCTA 2024 - 16th International Conference on Neural Computation Theory and Applications

578

The lines of short routes, if possible, run as straight as

possible, especially if they are also in the line of sight.

If the line of sight of the short route endpoint is

obscured by visual obstructions, the route may

deviate greatly from the straight line and detour along

the visible road section.

▪ Fragmentation of Observed Data.

The environment, including both artificial and natural

signs, is often underrepresented, contradictory, and

ambiguous; the boundaries, color, and texture are not

perceived clearly. The composition of objects

selected by a person changes over time, and when the

same objects are selected again, their fragments are

added and removed, and new combinations of

fragments are generated (LaPointe, Lupianez, and

Milliken 2013). In the observed environment,

fragments are selected that belong to one or several

objects, for example, not the road sign itself but some

part thereof, being jointly selected and even

combined with the adjacent lawn.

The success of the configuration options,

fragmentation, and recombination of blocks and

structural code can be determined by the statistics of

contradictions between a single found element and

the correspondence among the existing several

elemental options found exactly according to the

specified parameters. Attribute transferring allows to

smooth out corners and to make modest changes and

additions, and filling large voids in the data

approximately (Ge et al. 2021a, 2021b).

When overfitting, trained topics overwhelm

untrained topics by searching the environment for

previously encountered objects, phenomena, and

events. But potentially there is a transfer of attributes

from the known to the unknown. A list of objects or

fragments can create thematic non-overlapping layers

in a location that form new structures, this gradually

leads to changes in fields from the attribute table

where the transfer of attributes changes, for example,

color from green to red, and a car to a tractor in the

same field. Versions of assembly of layers are

possible. In such cases, a block combining elements

into a common or a consolidated block of mixed

elements or fragments from one or different sources

can be combined into a common whole with other

parts.

▪ Multiscale Spatial Code.

The nature of spatial perception can be described in

terms of topological entities, with visual form

primitives serving as key geometric invariants (Chen

2005). A multiscale spatial code is also present in the

brain, which allows external stimuli to be represented

with varying degrees of refinement, both in

generalization and in detail (Bellmund et al. 2018).

In problems of detecting visual changes,

researchers have shown that objects embedded in a

contextually heterogeneous scene tend to be detected

faster than objects embedded in a contextually

homogeneous scene. consisting of types of objects

and their probable location (LaPointe, Lupianez, and

Milliken 2013). The relative positions of points,

fragments, and objects form the structural code of the

route. The structural code is determined by the

relative location of commensurate objects and, in

general, does not change when objects are replaced

with their counterparts. Searching for objects with the

same attributes results in finding objects with the

same structural code.

The structure of a composite block of fragments

is a set of fragments and links between them. Such a

structural code of a composite block can be

transferred from one block to another. The structural

code for a set of objects is determined by the presence

or absence of boundaries adjacent to each other or to

the “background.” The structural code can change

when fragment blocks move relative to each other.

After a search for additional data in a collection of

fragments and blocks, an incomplete object can be

supplemented with missing data.

The structural code, like a route description, can

be used as a geometrical relative location of a set of

objects and as an invariant for describing motion.

Object attributes and structure codes often become

similar if they have been neighbors for a long time.

The difference between the existing and predicted

fragments can be calculated at the same location.

Tobler (Tobler 1970) assumes the dependence of

some attributes of objects that are close to each other.

Neighboring artificial objects may have a similar

structural code. If a group of objects is located in the

same area, attributes of objects are useful for restoring

the missing attributes of another object from the same

class. If objects of the same class are located in sight,

missing attributes can be added based on the available

attributes of other objects.

Search operations make it possible to find

intermediate points along a route. With redundant but

varied options, the lack of information is

compensated by additions found during the search.

▪ Signs and Dynamic Elements Along a Route.

In choosing a route, the goal is usually not a single

endpoint of interest but to obtain a variety of

“vacation” impressions, which is achieved by

selecting the entire set of points to visit—both final

and intermediate. Some of these POIs chosen by a

traveler may relate to their favorite hobby (places for

sports activities, cultural attractions, etc.), some POIs

are directly related to everyday needs, and some are

Spatial Learning and Overﬁtting in Visual Recognition and Route Planning Tasks

579

explicitly aimed at gaining new experiences (see

Figure 2).

Figure 2: Route signs in Stein am Rhein.

A sign is a material object or a pointer, which is

identified by many people in the environment, but can

have different imaginative content for any person.

However, for standard logistics tasks, identically

perceived signs are used. When exchanging data, a

sign usually represents a small amount of data and is

spatially localized. Signs along a route replace the

immediate appearance of a predicted event or external

information.

Points of interest often serve as bookmarks and

reference points in navigational applications. These

are required for a variety of cases of access to

information sources. Similar data from the external

world with supplements in the form of search results

restore the same initial conditions for collecting and

processing information. Being small in volume, POIs

allow quick recovery of information with the relevant

set of parameters, which was prepared earlier. Such

POIs often continue to be observed for a long time,

maintaining the sample selection stability when

searching.

Orientation information has the property of

maneuverability. When included in the current data

set together with other observed data, its fragments,

links, attributes, and structural codes are transferred.

Incoming data is modified by adding extra-large

fragments or reference points or by filling in missing

data using fragments. Expanding the currently

collected data using a search is also possible when

accessing extended data, with orientation information

acting as an instance of some class, or if there are

dependent or identical parts in the observed or found

elements.

▪ Choosing with Insufficient Information

If there is no fixed endpoint, the route can be very

tortuous. When fixing the start and end points of a

route, its intermediate points are movable and depend

on the selection and external events. According to

some signs, that are known and equally understood by

many people, it is possible to determine the type and

intensity of traffic at certain hours in a certain

direction.

Uncertainty in choosing a route can occur when

there is a lack or excess of information, or when it is

impossible to determine the importance of the

observed data. Choosing between two insignificant

options or avoiding a choice may be based on

intermediate surveying, or affected by external

influence, or accompanied by withdrawal from the

situation. Failure to pay attention to any POIs when

changing the thematic setting may be a result of

fatigue or overfitting.

Untrained – a deviating path taking into account

POIs, trained by teacher – the most well-trodden

paths – mostly based on the decisions of others,

taking into account the variations and, trained without

teacher – shortest paths.

Let the path consist of four parts: 1) planned

activities, 2) relaxation with inactive rest without

almost external events, c) visually attractive

contemplation or observation without temporary

haste for what is interesting and d) active spontaneous

entertainment here and now.

▪ Unexpected Events or New Road Signs

For some, perhaps a long, time synchronicity is

displayed in different structures and processes that

start at the same time. The synchronicity of objects

can be either temporal, as for a series of objects, or

spatial, as for a set of objects that are selected by a

person simultaneously (Ort and Olivers 2020). Even

if the geometric distance is fixed, due to weather and

other unexpected factors, traveling time from one city

to another may be variable (Neumann 2017). Short-

distance travel is associated with risk, when this

movement is carried out to link initially unrelated

segments of the route path that are close to each other

in the task of quick transition. It is worth noting that

often many short sections of the route can only be

traversed on foot.

Unexpected events, obstacles or new road signs

can significantly change both the route itself and the

set of intermediate points of interest. A small gap in

the planned route that arises due to a lack of

information regarding objects along the route is

closed under the assumption of the similarity of these

unknown objects with previously encountered ones.

Often, difficulties arise when opting for a

particular route; besides, natural and artificial

obstacles can affect the already planned path in the

most unpredictable ways. In such cases, where a

person is not able to generate a route model due to

unpredictable events or a mismatch with

expectations, they can always simplify the choice

model to a series of one-time choices (existing “here

and now”); however, in each new moment, the person

will again need to make next choice.

In addition to the existing “points of interest” on

a map of the area, a traveler can find new “points of

NCTA 2024 - 16th International Conference on Neural Computation Theory and Applications

580

attraction” that unexpectedly invite attention. They

can become intermediate points on the route.

The scale of the planned route determines the

features of its formation. When optimizing a route,

considering points of interest, one can draw an

isoline, connecting points with the same levels of

interest. For more flexibility in working with data,

one can use a dynamic segmentation of events along

the route. During the formation of the preliminary

route, it is assumed that both a choice has been made

and a class of objects to be approached has been

selected or the routes are offered to all known objects

within acceptable limits. During the search, objects

located in the vicinity of the route track can

supplement the set of criteria for route selection. A

not-fixed route endpoint increases the significance of

pointers located near the starting point.

The presence of a distinguishable choice between

two options does not imply the presence of a

predisposition to one of them. In this situation,

preparation is required for the solution: surveying in

situation with clarification of route details, external

interference to the situation, outside signs, and so on.

An important reason for the change or complete

cancellation of a route may be the discrepancy

between expected and real events.

3 MATERIALS AND METHODS

3.1 Materials

▪ Pigeon Flights

Spatial processing of GPS pigeon tracks was

performed based on data on bird flights in flocks over

the combined terrain, published in open repositories:

Dryad Digital Repository

https://datadryad.org/resource/doi:10.5

061/dryad.f9n8t, where flocks flew near the

seashore and Movebank Data Repository

https://doi.org/10.5441/001/1.33159h1),

where flocks flew near the foothills. The distance

between the points of departure and destination was

about 10 km. The distance in the coordinate sequence

of pigeon GPS tracks was about 3-6 meters.

Measurements of coordinates were taken 4–5 times

per second. The number of pigeons in each flock was

4-8 birds.

▪ Remote Sensing Data

Spatial remote sensing data were obtained from open

sources, such as Sentinel2 data hub

(https://www.sentinel-hub.com).

3.2 Processing Steps and Metrics

When analyzing the movement trajectories of birds,

our processing consisted of the following steps:

▪ Creating a new project in QGIS

(http://qgis.org) and uploading data

about trajectories and terrain;

▪ Identification of local key objects based on

remote sensing data, including selection of

contours of significant extended objects;

▪ Object recognition using neural networks;

▪ Calculation of the main indicators of tracks for

different degrees of learning;

▪ Calculation of track metrics.

When analyzing remote sensing data, we

accomplished data recognition in different versions of

networks:

▪ Non-specialized in remote sensing recognition,

but able to recognize images of a different type,

▪ Trained for remote sensing recognition,

▪ Trained, but with modified recognition

parameters - conditionally untrained.

Calculation of recognition metrics F1-Score_obj

(Lipton, Elkan, and Naryanaswamy 2014). To

calculate the F1-Score values, we used

sklearn.metrics module from open source machine

learning library Scikit-learn (https://scikit-

learn.org/stable/modules/generated/skle

arn.metrics.f1_score.html).

3.3 Applications

Spatial analysis was performed using QGIS

applications Extracts contour lines and Heatmap, and

LF Tools (https://github.com/LEOXINGU/

lftools/wiki/LF-Tools-for-QGIS).

The recognized remote sensing materials was

obtained using Mapflow AI platform

(https://github.com/Geoalert/mapflow-

qgis, http://mapflow.ai), which provides

geoinformation pipelines for recognizing objects

based on remote sensing data, such as buildings,

roads, fields, forests, etc. using various neural

network models. To improve the quality of

recognition, settings are specified for pre- and post-

processing of the results. Such settings provide a

fitting variation in recognition depending on sources

(aerial/satellite) and specifics of the recognized

classes (density of buildings, urban or forest

vegetation, etc.). Mapflow's recognition capabilities

can be used even to segment trees by crown type and

houses by height (see Figure 3).

Spatial Learning and Overﬁtting in Visual Recognition and Route Planning Tasks

581

Figure 3: Segmentation of trees by their crown types.

4 RESULTS

4.1 Analysis of Pigeon Flights

During the calculations, comparisons were made for

untrained, trained and overfitting cases of bird flights,

and it was found how pigeons fly without turning off

the path, depending on the types of flight and degree

of training.

Calculations were made for middle distance

tracks, where all distances between the points of

departure and destination of the flock of pigeons were

about 10 km. In GPS measurements were taken 4–5

times per second, distances between points of pigeon

GPS tracks was about 3-6 meters. The number of

pigeons in each group was 4-8 birds.

The following indicators were calculated:

▪ Straightness index (ratio of flight length in

target direction to total flight length);

▪ Flight around newly found POIs (percentage of

cases of flying over the buffer zone of the flight

to explore new POIs).

Figure 4: Typical flights: (a) Surveying flight; (b) Trained

flight; (c) Trained flight.

Typical pigeon flights are shown in the Figure 4.

The calculation took into account the pigeon's

level of training, in accordance with the number of

flights performed along a given route. It was believed

that the 1st-3rd flight is an untrained or surveying

pigeon; the 3rd–7th flight is a trained pigeon; above

the 7th flight is an overfitting pigeon.

The calculated results of pigeons’ flight

efficiency for surveying flights (the bird does not yet

know the way), trained flights (the bird knows the

way, but is distracted by external factors), and

overfitting flights (the bird flies as directed as

possible towards its target) are shown in the Table 1.

Table 1: Pigeon’s flight efficiency.

Evaluation

Surveying

flights

Trained

flights

Overfitting

flights

Straightness

index

0.09

0.64

0.89

Flight around

POIs

64 %

27%

11%

4.2 Efficiency of Neural Networks

Examples of building recognition by untrained,

trained, and overfitting neural networks (NN) are

shown in the Figure 5. It is noticeable that the

untrained neural network tries to find buildings based

on any minimal features, the trained neural network

detects buildings with high efficiency, and the

overfitting neural network correctly recognizes

buildings in images of a familiar type but at the same

time makes a large number of errors on unfamiliar

textures.

Figure 5: Buildings recognition: (a) untrained NN; (b)

trained NN (c) overfitting NN.

The calculated results of F1-Score are shown in

the Table 2.

Table 2: Untrained, trained, and overfitting NN.

Evaluation

Untrained NN

Trained

Overfitting

Number of

training samples

(1 x 1 km

)

Recognized

objects in

training samples

265

854

1430

F1-score

0.17

0.87

0.52

NCTA 2024 - 16th International Conference on Neural Computation Theory and Applications

582

5 CONCLUSIONS

Spatial learning in perception and navigation are

essential skills in a changing environment. Both in AI

and in nature, there is a problem of “overfitting” when

a bird accustomed to the same route fails to notice

new places to forage, or an artificial neural network

begins to detect buildings in the ripples of the ocean.

The “challenge” of overfitting makes it difficult to

obtain new information and to find optimal solutions.

Special attention is paid to the problems of overfitting

when detailed adherence to previously acquired

behavioral patterns leads to a decrease in efficiency

and the accumulation of systematic errors.

Additionally, due to overfitting, the ability to make

optimal decisions in the presence of significant

changes in the environment is reduced.

Our work systematizes general issues related to

spatial data processing. We examine the problem of

learning and retraining in spatial cognition and

navigational behavior in categories: birds’ navigation

behavior and remote sensing recognition with neural

networks and demonstrates techniques for solving the

problem of overfitting. It can be helpful in various

industry applications, including tracking changes in

animal migrations in conditions of climate change,

creating smart interactive tourist routes and adapting

infrastructure for tourism, and preparing new neural

network models for recognizing spatial data.

REFERENCES

Alharbi, A. A. (2024). “Classification Performance

Analysis of Decision Tree-Based Algorithms with

Noisy Class Variable.” Discrete Dynamics in Nature

and Society, pp. 1–10.

Bashir, D. et al. (2020). “An Information-Theoretic

Perspective on Overfitting and Underfitting.” AI 2020,

pp. 347–58.

Bellmund, J. L. S. et al. (2018). “Navigating Cognition:

Spatial Codes for Human Thinking.” Science.

362(6415): eaat6766.

Benedetti, P., Femminella, M., Reali, G. 2022. “Mixed-

Sized Biomedical Image Segmentation Based on U-Net

Architectures.” Applied Sciences. 13(1): 329.

Bengar, J. Z. et al. (2021). “When Deep Learners Change

Their Mind: Learning Dynamics for Active Learning.”

CAIP 2021, pp. 403–13.

Blaser, N. et al. (2013). “Testing Cognitive Navigation in

Unknown Territories: Homing Pigeons Choose

Different Targets.” Journal of Experimental Biology.

216(16), pp. 3123–31.

Chen, L. (2005). “The Topological Approach to Perceptual

Organization.” Visual Cognition. 12(4), pp. 553–637.

Dietterich, T. G. (2002). “Machine Learning for Sequential

Data: A Review.” SSPR/SPR, pp. 15–30.

Foster, J., Schoepf, S., Brintrup, A. (2024). “Fast Machine

Unlearning without Retraining through Selective

Synaptic Dampening.” AAAI. 38(11), pp. 12043–51.

Freksa, C. et al. (2017). “Spatial Problem Solving in Spatial

Structures.” MIWAI 2017, pp. 18–29.

Galvani, A., Zaleshina, M., Zaleshin, A. (2021). “Cognitive

Geography. Space Reflected in the Mind.” Hidden

Geographies, pp. 41–52.

Ge, Y. et al. (2021). “Zero-Shot Synthesis with Group-

Supervised Learning.” ICLR 2021, pp. 1–16.

Ge, Y. et al. (2021). “A Peek Into the Reasoning of Neural

Networks: Interpreting with Structural Visual

Concepts.” CVPR, pp. 2195–2204.

Hopkins, M. et al. (2024). “Realizable Learning Is All You

Need.” TheoretiCS.

Kim, T., Kim, S., Bengio, Y. (2021). “Visual Concept

Reasoning Networks.” AAAI. 35(9), pp. 8172–80.

Kong, X. et al. (2018). “Big Trajectory Data: A Survey of

Applications and Services.” IEEE Access.

LaPointe, M.R.P., Lupianez, J., Milliken B. (2013).

“Context Congruency Effects in Change Detection:

Opposing Effects on Detection and Identification.”

Visual Cognition. 21(1), pp. 99–122.

Lee, J. et al. (2022). “A Pixel-Level Coarse-to-Fine Image

Segmentation Labelling Algorithm.” Scientific Reports.

12(1): 8672.

Lipton, Z.C et al. (2014). “Optimal Thresholding of

Classifiers to Maximize F1 Measure.” Mach Learn

Knowl Discov Databases, pp. 225–39.

Neumann, T. (2017). “Vessels Route Planning Problem

with Uncertain Data.” TransNav, the International

Journal on Marine Navigation and Safety of Sea

Transportation.

Ort, E,, Olivers, C. (2020). “The Capacity of Multiple-

Target Search.” Visual Cognition, pp. 1–26.

Santos, C.D. et al. (2014). “Temporal and Contextual

Consistency of Leadership in Homing Pigeon Flocks.”

PLoS ONE. 9(7): e102771.

Schiffner, I. et al. (2018). “Behavioural Traits of Individual

Homing Pigeons, Columba Livia f. Domestica, in Their

Homing Flights.” PLoS ONE.

Siddique, N. et al. (2017). “A Review of the Relationship

between Novelty, Intrinsic Motivation and

Reinforcement Learning.” Journal of Behavioral

Robotics. 8(1), pp. 58–69.

Tobler, W R. (1970). “A Computer Movie Simulating

Urban Growth in the Detroit Region.” Science. 46, pp.

234–40.

Tversky, A., Kahneman, D. (1974). “Judgment under

Uncertainty: Heuristics and Biases.” Science.

185(4157), pp. 1124–31.

Wang, Y. et al. (2022). “An Improved Deeplabv3+

Semantic Segmentation Algorithm with Multiple Loss

Constraints.” PLOS ONE. 17(1): e0261582.

Zhao, W., Alwidian, S., Mahmoud, Q.H. (2022).

“Adversarial Training Methods for Deep Learning: A

Systematic Review.” Algorithms. 15(8): 283.

Spatial Learning and Overﬁtting in Visual Recognition and Route Planning Tasks

583