Machine Learning Techniques and the Existence of Variant Processes

in Humans Declarative Memory

Alex Frid

, Hananel Hazan

, Ester Koilis

, Larry M. Manevitz

3,5

, Maayan Merhav

and Gal Star

Edmond J. Safra Brain Research Center, University of Haifa, Haifa, Israel

Network Biology Research Laboratory, Technion, Haifa, Israel

Computer Science Department, University of Haifa, Haifa, Israel

German Center for Neurodegenerative Diseases (DZNE), Magdeburg, Germany

Center of Information and Neural Networks, National Institute of Information and Communications Technology,

Suita, Osaka, Japan

Keywords: Machine Learning, Classification, functional Magnetic Resonance Imaging (fMRI), Feature Selection,

Support Vector Machines, Radial Basis Function Kernel, Declarative Memory, Information Biomarkers.

Abstract: This work uses supervised machine learning methods over fMRI brain scans to establish the existence of

two different encoding procedures for human declarative memory. Declarative knowledge refers to the

memory for facts and events and initially depends on the hippocampus. Recent studies which used patients

with hippocampal lesions and neuroimaging data, suggested the existence of an alternative process to form

declarative memories. This process is triggered by learning mechanism called "Fast Mapping (FM)", as

opposed to the 'standard' "Explicit Encoding (EE)" learning procedure. The present work gives a clear

biomarker on the existence of two distinct encoding procedures as we can accurately predict which of the

processes is being used directly from voxel activity in fMRI scans. The scans are taken during retrieval of

information wherein the tasks are identical regardless of which procedure was used for acquisition and by

that reflect conclusive prediction. This is an identification of a more subtle cognitive task than direct

perceptual cognitive tasks as it requires some encoding and processing in the brain.

1 INTRODUCTION

Human declarative memory is defined as the

conscious information recollection of facts and

events (Squire, 1992). Under the "standard model"

theory for adult declarative memory systems, novel

information is encoded explicitly into the memory

using, amongst other brain parts, the hippocampus

(McClelland et al., 1995). This standard,

hippocampal dependant memory is acquired through

intentional "Explicit Encoding (EE)" procedure. The

encoded information is then slowly transferred from

the hippocampus to the neo-cortex where it becomes

permanently stored (Squire and Alvarez, 1995;

Frankland and Bontempi, 2005). Overtime, the

initially hippocampal dependant memories become

independent of the hippocampus. It has been

suggested that this re-organization process is done

during sleep (Gais et al., 2007).

Amongst toddlers, the process of rapid language

acquisition occurs prior to the full development of

the hippocampus (Bauer, 2008; Uematsu et al.,

2012). Moreover, some evidence from hippocampal

injured subjects demonstrated an ability to acquire

information which seems to have declarative-like

characteristics despite severe damages in the

hippocampus (Sharon et al., 2011; Merhav et al.,

2014) and so must involve a different brain network

than the one engaged by "EE". This alternative

learning mechanism is called "Fast Mapping (FM)".

It is unknown if the memory representations

following FM undergo consolidation processes,

similar to memories gained through EE. However,

since it was shown that patients with hippocampal

damages as well as healthy controls could learn and

store information acquired via FM for a week

(Sharon et al., 2011; Merhav et al., 2014), the

scheme used to explain memory consolidation of

other declarative memories cannot be applied for

FM in a straightforward manner.

It remains somewhat controversial as to whether

the FM is available for acquisition of words among

amnesic patients (Warren and Duff, 2014). Proving

that FM methods are mostly based on brain

114

Frid, A., Hazan, H., Koilis, E., Manevitz, L., Merhav, M. and Star, G..

Machine Learning Techniques and the Existence of Variant Processes in Humans Declarative Memory.

In Proceedings of the 7th International Joint Conference on Computational Intelligence (IJCCI 2015) - Volume 3: NCTA, pages 114-121

ISBN: 978-989-758-157-1

structures outside of the hippocampus area opens

possibility for therapeutic approach for people with

damages in these areas.

In this work we aim to demonstrate the

distinctiveness of brain systems, which support EE

and FM memory process, by extracting activity

patterns directly from brain data. Functional

magnetic resonance imaging (fMRI) captures

information from thousands of different localities

(voxels) of the brain simultaneously. Multivariate

pattern analysis approach (MVPA) (Norman et al.,

2006) utilizes these activities by looking for changes

in BOLD signal across different voxels. Different

methods can be used for analysis on such complex

data depending on the question of study (retrieval or

decoding stimuli, mental states, behaviours and

other variables of interest). A growing number of

studies (Mitchell et al. 2008; Kriegeskorte et al.,

2006; Nawa and Ando, 2014; Atir-Sharon et al.,

2015) shows ability in using machine learning

methods for analysis of neuroimaging data.

Nevertheless, the feasibility to achieve successful

results using machine learning on fMRI multivariate

data is not trivial and relies on the sensitive choice

of features to be considered in the analysis.

2 RELATED WORK

The mechanism of FM was examined among healthy

individuals (Gilboa et al., 2011; Atir-Sharon et al.,

2015). It was shown that two learning mechanisms,

EE and FM, can be discriminated from fMRI data

during memory acquisition using machine learning

based classifier. In addition, memories acquired

during scanning were tested for recollection success

later, outside the fMRI machine. Successful

accuracy results were achieved when identifying

scans corresponding to successful and unsuccessful

recollection within EE group and within FM group,

for each participant separately and cross-participant.

However, the different nature of the procedures

used for acquisition of information (EE and FM),

does not allow for complete control over the task

with regard to the behavioural experience.

Therefore, the possibility remained that the

successful classification obtained in the experiment

is a result of differences in the acquisition

procedures and not in the learning mechanisms.

To overcome this limitation, in another study

(Merhav et al., 2015), the neural correlates of FM

and EE were explored during a retrieval procedure,

designed to be identical for both mechanisms. In

addition, the study was focused on overnight re-

organization of memory representations, following

both EE and FM. Findings suggested that, despite

the identical retrieval tasks, memories that were

gained through FM induced distinct neural

substrates from those involved EE (Merhav et al.,

2015). While retrieval of data learned through EE

engaged the expected hippocampal and vmPFC

related network, retrieval of information acquired

through FM immediately engaged an ATL related

network, typically supporting well-established

semantic knowledge. In addition, analysis of

neuroimaging data associated with EE showed the

expected overnight changes in network connectivity

where for FM minimal overnight changes were

presented. The analysis was performed by a

multivariate technique of Spatiotemporal Partial

Least Squares (PLS), helping to identify assemblies

of brain regions that co-vary together.

3 CURRENT STUDY

In this study, fMRI brain data was captured during

retrieval of memories, acquired through either EE or

FM. The goal is to provide a biomarker directly

from these fMRI scans using machine learning

methods. Such classification ability based on the

neural activity data gives strong evidence for the

existence of distinct neural processes associated with

EE and FM.

Multivariate classification is performed on fMRI

features obtained during memory recollection, where

tasks performed by the participants are identical for

EE and FM. We also perform classification to

explore re-organization processes following both

learning mechanisms. Classification was performed

over brain scans which were acquired either 30

minutes before scanning (recent memory) or a day

before scanning (remote memory).

Regarding the distinction between the two

memory processes during recollection, we address

two questions:

1. Is it possible to distinguish between the two

learning modes (i.e. EE and FM) based on

neural activity information collected during

the recollection of memories?

2. Is it possible to distinguish between items

learned recently and remotely?

Machine Learning Techniques and the Existence of Variant Processes in Humans Declarative Memory

115

4 EXPERIMENT PROCEDURE

4.1 Participants

The experiment, full details of which can be seen in

Merhav et al. (Merhav et al., 2015), was conducted

in Rotman Research Institute at Baycrest, Canada.

Here we mention the salient points.

32 participants (20 females) were recruited and

randomly assigned to one of the two groups (EE or

FM). All participants were English native speakers,

right-handed and had no history of neurological or

psychiatric disorders and no learning disabilities. A

written informed consent was obtained according to

Baycrest’s Research Ethics Board’s guidelines.

Gender and age distributions (10 females in each

group) were similar in the FM and in the EE groups,

respectively. The two groups also did not differ on

the number of years of education, I.Q. estimates and

WMS-III Verbal Paired Associates retention.

4.2 Experiment Paradigm & Procedure

32 healthy adult participants (20 females) were

randomly assigned of one to two groups (EE or FM).

On day 1 the participants learned 50 new unfamiliar

picture-word associations. On day 2 (24 hours later)

they learned another set of 50 new picture-word

associations. A retrieval memory test for all the 100

new picture-word associations took place 30 minutes

after the acquisition of a second set of associations.

During the retrieval, brain activity was scanned

(Figure 1A). Therefore, the participants were tested

on both recently and remotely encoded information.

The two learning tasks (EE / FM) were designed

differently due to different nature of both learning

procedures (Figure 1B).

The retrieval task was designed as an event

related fMRI experiment in which memory for all

100 items was assessed via an associative four-

alternative forced choice recognition task. The

retrieval procedure was identical for EE and FM as it

was performed inside the scanner (Figure 1C). Each

retrieval trial of an item was 12.5 seconds long and

contained the following intervals: blank screen (1

sec), target label as text and auditory input (1.5 sec),

4 choice pictures appeared on screen, below the

target label (2.5 sec), the word "choose" appeared

onscreen and participants had to respond by

selecting the appropriate key (5 sec), confidence

rating (2.5 sec).

The experiment was designed intentionally to

have participants perform either EE or FM, rather

than perform both EE and FM tasks. It was

important that learning through FM will be implicit

and unintentional, so participants should not know

that the task is a mnemonic task (i.e., requires

memory). However, in EE, participants are explicitly

asked to remember the name of the item.

4.3 Data Acquisition & Pre-processing

The participants were scanned using the Siemens

Trio 3 T scanner, at Baycrest Institute. They

acquired T2*-weighted images, covering the whole

brain using an echo-planar imaging (EPI) sequence

of 50 slices, with repetition time (TR) of 2500 ms,

echo time (TE) of 27 ms, 64 × 64 matrix, slice

thickness of 3.5 mm and a field of view (FOV) of

200 mm. The procedure was designed as an event

related fMRI study.

Figure 1: (A) The experiment structure. (B) Examples of

acquisition through FM (left) and through EE (right). (C)

Retrieval test design which took place inside the fMRI

scanner.

The pre-processing steps included conversion to

4-dimensional AFNI format (Cox, 1996), slice

timing correction using the first slice as a reference

(Figure 2A), movement correction for unintended

head motions and spatial smoothing with 6mm

FWHM Gaussian kernel to increase signal-to-noise

ratio (Figure 2B). Finally, individual participant's

NCTA 2015 - 7th International Conference on Neural Computation Theory and Applications

116

data was converted to a standard coordinate system

(Talairach) to allow data analysis across individuals.

The scanning of each participant was done

during four consequent runs creating a joint dataset

out of four time-series datasets with approximately

150 data volumes each of size 109x91x91, resulted

as a dataset with approximately 600 data volumes.

Therefore, each data volume (data point) contained

1490580 different voxels. We demonstrate the

structure of the collected data in Figure 3.

Figure 2: Examples for pre-processing steps on fMRI data.

(A) Correction of individual's hemodynamic responses

slices acquired aligned to the exact same time (Sladky et

al., 2011). (B) Performance of spatial smoothing on fMRI

volume taken from single participant.

Figure 3: 4-Dimensional structure of AFNI format BRIK

(Cox, 1996) file including 3-dimensional dataset over time

sequence.

5 METHODS

The data points used for analysis were constructed

using scan data obtained for TR=2. This temporal

cut was selected after performing pre-test

classification as suggested in Atir-Sharon et al. work

(Atir-Sharon et al., 2015), taking into consideration

the accordance to the expected HRF response.

We performed further pre-processing over the

time-series data. At first, all non-brain voxels were

removed using a mask. This was done by selecting

voxels from the fMRI dataset that correspond to

non-zero elements in the mask (creating data points

of approximately 200,000 voxels). Afterwards,

linear detrending was performed on each

participant's data set and for each run separately.

Then, normalization over all scans was

conducted. The normalization was done voxel-wise

using z-score for each participant separately. In our

case, the combined dataset involved scans from

different groups and participants taken from

different distributions. Therefore, transformation of

features from different scales to a single scale, with

consideration to the original distributions, was

needed. The z-score method considers the different

distribution characteristics of every group (Wiesen,

2006), hence, it was chosen as the normalization

procedure. The z-score formula is presented in (1),

where z-val is the new z-scored value, f-val is the

original feature value and (μ, σ) are the mean and

standard deviation values:

z-val = (f-val – μ) / σ (1)

For the mean and standard deviation computation in

the z-score equation, several assignments were

tested: (i) from all scans in the dataset; (ii) from

individual participants' scans and (iii) from the

distribution of scans marked as control (baseline) in

the training set. Best classification results were

achieved by using the mean and standard deviation

computed from the distribution of baseline scans

(option (iii)).

Each volume was represented as an individual

data point in the dataset (i.e. each voxel was

considered as a feature). Since the amount of scans

from EE and FM groups was not equal, counter-

balancing of the dataset was performed. This was

done by randomly sampling data points from the

smaller group. This method was applied only on the

training set. Otherwise, more weight would have

been given to prediction accuracies of duplicated

data points against weight of accuracies for data

points that were not duplicated. Therefore, testing

set was left untouched.

Machine learning classification techniques were

used for data analysis. Considering the high

dimensionality of data used in the current study,

feature selection procedure was performed in order

to reduce the number of features used for

multivariate classification analysis. There are several

generic methods for selecting informative features.

Machine Learning Techniques and the Existence of Variant Processes in Humans Declarative Memory

117

We aimed to select the features that best

discriminate between conditions based on their

activation values. It was achieved by ranking the

importance of each feature according to the ANOVA

F-score value obtained for between-group (EE vs.

FM) comparisons.

To find the optimal subset of features for

analysis, we performed exhaustive search for

different sizes of features sets starting from 10

features to full brain features in exponential manner.

Finally, the top 1000 features with highest F-scores

were selected. This relatively large number of

features was chosen to take advantage of inclusion

of weakly informative voxels which can contribute

to an increase in classification rates (Gonzalez-

Castillo et al., 2012). In Figure 4, we illustrate the

extracted features in the form of a brain map. In this

example, we display in red selected subset of

features for recollection (correct vs. Incorrect)

classification. This was performed on individual's

fMRI data that belongs to the FM group. Note, that

not all the selected features can be depicted in a

single brain map, but it can be seen that they

concentrated in a specific areas.

A cross-validation classification scheme using

Support Vector Machine classifier (Vapnik, 1998)

with RBF (Radial Basis Function) kernel (Vert,

Tsuda and Schölkopf, 2004) was applied to the

selected features.

Parameters that are not learnt directly within

estimators can be set by searching a parameter space

for the best cross-validation score. Grid search for C

and gamma parameters was performed in the ranges

of 2

-5

to 2

and 2

-15

to 2

respectively. Grid search

was executed before training on a training portion of

the dataset to achieve increase in accuracy rates. A

pseudo-code for the performed grid search is

presented in Figure 5. In all runs parameters C and

gamma were set to 1 and 2

-3

respectively.

Figure 4: Brain map displaying features selected for

classification analysis of FM recollection (correct vs.

Incorrect).

In cases where the testing set consisted of scans that

for c in [2

-5

, 2

-3

,...,2

for g in [2

-15

, 2

-13

, ..., 2

for train, test in partition:

model = svm_train(train, c, g)

score = svm_predict(test, model)

cv_list.insert (score)

scores_list.insert(mean(cv_list),c,g)

print max(scores_list)

Figure 5: Pseudo-code for grid search procedure.

were taken from one group only (i.e. all scans were

EE or all scans were FM), a decision making

function was applied. We used majority voting

method as a decision making function, defined as

follows: if the majority of the scans were rated

correctly per participant, the accuracy was set to 1,

otherwise, the accuracy was set to 0.

The software used for the classification was

developed using Python programming language and

based on LibSVM (Chang and Lin, 2011) and

PyMVPA software packages (Hanke et al., 2009). In

Figure 6 we present a complete analysis flow

diagram including all the relevant pre-processing

and processing stages.

6 RESULTS

6.1 Memory Performance

In the information retrieval test, correct response

rates for the recent and for the remote associations

were significantly above chance (binomial tests, p <

0.0001, for both times-of-acquisition, in both

learning groups). Overall, participants from the FM

group were less successful in retrieval, compared to

those from the EE group, in both the recent and the

remote conditions (F(1,30) = 12.2, p < 0.005).

In both groups, recent items were better

recognized than remotely presented items (F(1) =

9.12, p = 0.005) with no significant interaction

between the time of acquisition and the learning

mode (F(1,15) = 0.334, p = 0.565).

6.2 Classification

First, we addressed the question of classifying scans

obtained during correct and incorrect recollection.

Using the proposed classification scheme, we

performed 4-fold (leave one run out) cross-

validation within participants. The mean values of

classification accuracy were close to the chance

level for both groups (EE and FM). We theorized the

NCTA 2015 - 7th International Conference on Neural Computation Theory and Applications

118

Figure 6: Schematic diagram of the steps performed for whole brain analysis procedure. It consist the following stages: (A)

The initial stage representing the neuroimaging data delivery. (B) The pre-processing stage. (C) Data reduction stage:

reducing data variability efficiently by feature selection. (D) Learning stage: performing multiple times by cross validation

procedure.

reason for that is the existence of two additional

different sub-groups, recent and remote word

acquisition, within each of the initial groups.

Therefore, we classified correct and incorrect scans

within each possibility: EE recent, EE remote, FM

recent, FM remote. For each possibility we chose

10% of all data points randomly as a testing set. The

rest of the data points were used for training. Then,

10-fold cross validation was performed. We report

the values for mean and standard deviation of

classification accuracy over 10 cross-validation folds

for EE in Table 1 and for FM in Table 2.

These results show that a trained classifier was

able to distinguish scans obtained during correct and

incorrect word recollection within each group. The

accuracy is higher for classification of scans for

words learned recently, rather than for words learned

remotely. Furthermore, the discriminating ability is

better within EE group rather than within FM group.

Next, we classified whether the process used for

information acquisition was EE or FM using only

scans from the successful recollection attempts in

the behavioural experiment. We chose randomly

10% from all the scans of all participants as a testing

set. The rest of the scans were used as a training set.

The values and standard deviations for classification

accuracy are presented in Table 3. The results show

that using the neuroimaging data from each one of

the participants for training, we could distinguish

between EE and FM scans very well.

Table 1: Correct vs. Incorrect classification within Explicit

Encoding (EE) using 10-fold cross validation.

Mean Accuracy Standard

Deviation

Recent

0.708 0.09

Remote

0.584 0.067

Table 2: Correct vs. Incorrect classification within Fast

Mapping (FM) using 10-fold cross validation.

Mean Accuracy Standard

Deviation

Recent

0.599

0.063

Remote

0.55

0.068

Table 3: EE vs. FM (using only correct recollection scans)

across participants using 16-fold cross-validation.

Testing set

selection method

Mean

Accuracy

Standard

deviation

Random selection

0.937

0.069

Leave one

participant out

0.638

0.07

These results raise the question of whether the

representation of all the participants in the training

set is crucial to the classification success. That is,

can a machine learning classifier, trained over the

collected data, can successfully distinguish which

label to assign to a new person scan, despite the fact

that the classifier has never seen data from this

participant. To answer this question, we performed a

leave-one-participant-out classification. This was

Machine Learning Techniques and the Existence of Variant Processes in Humans Declarative Memory

119

done across all 16 participants in a cross-validation

manner (leave one participant out). Note that per

iteration, the scans in the testing set are all EE or all

FM. Therefore, we were able to use the majority

voting method for this analysis. The results averaged

across all participants presented in Table 3.

7 DISCUSSION & CONCLUSIONS

In this work, we showed that it is possible to identify

correct and incorrect recollection of memories

acquired through two learning mechanisms: either

Explicit Encoding (EE) or Fast Mapping (FM)

directly from neuroimaging data using machine

learning techniques. The findings suggest that it is

easier to identify recollection success and failure for

information acquired recently rather than for

information after a period of time through EE

mechanism. It may indicate that the newly gained

information, acquired through EE, has started to take

part in consolidation process. At the same time, no

significant change between recollection results of

recent and remote acquisition was seen within the

FM mechanism. This may indicate that FM does not

engage consolidation processes. Further

classification experiments are required to reach a

more general conclusion.

The current results provide additional evidence

for the existence of two memory formation

processes by successfully classifying scans of

correct retrievals following EE and FM. Note that

the classification results for scans taken from an

individual’s data, which were not used previously

for training, were still significant (although less

accurate when training data from a subject were

included). These findings suggest that associative

learning through FM employs alternative neural

pathways to acquire declarative knowledge, which

bypasses the dominant hippocampal-vmPFC axis.

This also indicates that the FM process is eligible for

therapeutic approach for people with hippocampal

brain injuries.

8 FUTURE WORK

Future work should include mapping of the brain

regions and extraction of functional networks

associated with all four group combinations, EE

recent, EE remote, FM recent and FM remote. A list

of possible implementation approaches includes

constructing brain maps using "searchlight"

techniques (Kriegeskorte et al., 2006).

In addition, future work should include brain

regions correlations tests during the retrieval of

memory through EE and through FM in recent and

in remote modes. Those correlations would provide

information regarding the involvement of the

hippocampus and vmPFC regions in the

consolidation processes. To achieve that, one may

use causality analysis techniques (Hu & Liang,

2012) to reveal the causality influences the brain

regions, which are involved with each learning

procedure, have on each other. This could help

reveal new information regarding the mechanism

involved in memory consolidation processes of FM.

ACKNOWLEDGEMENTS

This work is part of the M.Sc thesis of Ms. Gal Star

at University of Haifa under the supervision of Prof.

Larry Manevitz at the Neuro-Computation

Laboratory at Caesarea Rothschild Institute (CRI),

Haifa, Israel.

The research is based on data gathered by

Rotman Research Institute at Baycrest, Toronto,

Canada. The examining of this data was suggested

by Dr. A. Gilboa and complements the work of

Merhav, Karni and Gilboa (Merhav et al., 2015).

The authors are listed in alphabetical order.

REFERENCES

Atir-Sharon, T., Gilboa, A., Hazan, H., Koilis, E. &

Manevitz, L. M. 2015. "Decoding the formation of

new semantics: MVPA investigation of rapid

neocortical plasticity during associative encoding

through Fast Mapping.". Neural Plasticity, vol. 2015,

Article ID 804385, 17 pages.

Bauer, P. J., 2008. "Toward a neuro-developmental

account of the development of declarative memory".

Dev Psychobiol, vol. 50, no. 1, pp. 19-31.

Chang, C. C. & Lin, C. J., 2011. "LIBSVM: a library for

support vector machines". ACM Transactions on

Intelligent Systems and Technology, available from:

<http://www.csie.ntu.edu.tw/~cjlin/libsvm>.

Cox, C., 1996. "AFNI: software for analysis and

visualization of functional magnetic resonance

images", Computers and Biomedial Research, vol. 29,

pp. 126-173.

Frankland, P. W., and Bontempi, B., 2005. "The

organization of recent and remote memories". Nature

Review: Neuroscience, vol. 6, pp. 119-130.

Gais, S., Albouy, G., Boly, M., Dang-Vu, T.T., Darsaud,

A., Desseilles, M., Rauchs, G., Schabus, M.,

Sterpenich, V., Vandewalle, G., Maquet, P., Peigneux,

P., 2007. "Sleep transforms the cerebral trace of

NCTA 2015 - 7th International Conference on Neural Computation Theory and Applications

120

declarative memories". Proceedings of the National

Academy of Sciences of the USA, vol. 104, no. 47, pp.

18778-18783.

Gilboa, A., Hazan, H., Koilis, E., Manevitz, L. and

Sharon, T, 2011. "Two memory systems: identifying

human memory encoding mechanisms from

psychological fMRI data via machine learning

techniques". Proceedings of the International Joint

Conference on Neural Networks (IJCNN), pp. 54.

Gonzalez-Castillo, J., Saad, Z.S., Handwerker, D.A., Inati,

S.J., Brenowitz, N., Bandettini, P.A., 2012. "Whole-

brain, time-locked activation with simple tasks

revealed using massive averaging and model-free

analysis". Proceedings of the National Academy of

Sciences, vol. 109, no. 14, pp. 5487-5492.

Hanke, M., Sederberg, P. B., Hanson, S. J., Haxby, J. V.,

and Pollmann, S., 2009. "PyMVPA: A python toolbox

for multivariate pattern analysis of fMRI data".

Neuroinformatics, vol. 7, no. 1, pp. 37-53.

Hu, S. & Liang, H., 2012. "Causality analysis of neural

connectivity: New tool and limitations of spectral

granger causality". Neurocomputing, vol. 76, no. 1, pp.

44-47.

Kriegeskorte, N., Goebel, R., and Bandettini, P., 2006.

"Information-based functional brain mapping".

Proceedings of National Academy of Science USA,

vol. 103, no. 10, pp. 3863-3868.

McClelland, L., McNaughton, B. L., and O'Reilly, R. C.,

1995. "Why there are complementary learning system

in the hippocampus and neo-cortex: insights from the

successes and failure of connectionist models of

learning and memory". Psychological Review, vol.

102, no. 3, pp. 419-457.

Merhav, M., Karni, A. and Gilboa, A., 2014. "Neocortical

catastrophic interference in healthy and amnesic

adults: A paradoxical matter of time". Hippocampus,

vol. 24, no. 12, pp. 1653-1662.

Merhav, M., Karni, A. and Gilboa A., 2015. "Not all

declarative memories are created equal: fast mapping

as a direct route to cortical declarative

representations". Neuroimage, vol. 117, pp. 80-92.

Mitchell, T., Shinkareva, S., Carlson, A., Chang, K. M.,

Malave, V. L., Mason, R. and Just M. A., 2008.

"Predicting human brain activity associated with the

meanings of nouns". Science, vol. 320, no. 5880, pp.

1191-1195.

Nawa, N. E. & Ando H., 2014. "Classification of self-

driven mental tasks from whole-brain activity

patterns". PLos One, vol. 9, no. 5, e97296.

Norman, K. A., Polyn, S. M., Detre, G. J. & Haxby, J. V.,

2006. "Beyond mind-reading: multi-voxel pattern

analysis of fMRI data". Trends in cognitive science,

vol. 10, no. 9, pp. 424-430.

Sharon. T., Moscovitch, M., and Gilboa, A., 2011. "Rapid

neocortical acquisition of long-tem arbitrary

associations independent of the hippocampus".

Proceedings of the National Academy of Science of the

USA, vol. 108, no. 3, pp. 1146-1151.

Sladky, R., Friston, K. J., Tröstl, J., Cunnington, R.,

Moser, E. & Windischberger, C., 2011. "Slice-timing

effects and their correction in functional MRI".

Neuroimage, vol. 58, no. 2, pp. 588-594.

Squire, L. R., 1992. "Declarative and non-declarative

memory: multiple brain systems supporting learning

and memory", Journal of Cognitive Neuroscience, vol.

4, no. 3, pp. 232-243.

Squire, L. R., and Alvarez, P., 1995. "Retrograde amnesia

and memory consolidation: a neurobiological

perspective". Current Opinion in Neurobiology, vol. 5,

no. 2, pp. 169-177.

Uematsu, A., Matsui, M., Tanaka, C., Takahashi, T.,

Noguchi, K., Suzuki, M. and Nishijo, H., 2012.

"Developmental trajectories of amygdale and

hippocampus from infancy to early adulthood in

healthy individuals". PLos One, vol. 7, no. 10, e46970.

Vapnik, V., 1998. "Statistical learning theory". New York,

NY: Wiley.

Vert, J. P, Tsuda, K. and Schölkopf, B., 2004. "A primer

on kernel methods". Kernel Methods in Computational

Biology.

Warren, D. E. and Duff, M. C., 2014. "Not so fast:

Hippocampal amnesia slow word learning despite

successful fast mapping". Hippocampus, vol. 24, no. 8,

pp. 920-933.

Wiesen J. P., 2006. "Benefits, Drawbacks, and Pitfalls of

z-Score Weighting". 30th Annual IPMAAC

Conference. Available at: "http://annex.ipacweb.org/

library/conf/06/wiesen.pdf" (27 Jun 2006).

Machine Learning Techniques and the Existence of Variant Processes in Humans Declarative Memory

121