full-FORCE: A target-based method for training recurrent networks

DePasquale, Brian; Cueva, Christopher J.; Rajan, Kanaka; Escola, G. Sean; Abbott, L. F.

doi:10.1371/journal.pone.0191527

Cited by 130 publications

(164 citation statements)

References 31 publications

Supporting

Mentioning

162

Contrasting

Order By: Relevance

“…It is also worth noting that other mechanisms for dimensionality expansion exist for recurrent networks and would be interesting to explore in future works. Examples include training the network with an explicit dimensionality term in the loss function or implicitly via a highly variable "target" network as in [26,12]. Indeed, we have used only a single, basic type of "vanilla RNN" network model, and extensions toward more complex models is important if we are to generalize our findings to other machine learning settings, and to make more confident predictions about the brain's circuits.…”

Section: Conclusion and Discussionmentioning

confidence: 99%

Dynamic compression and expansion in a classifying recurrent network

Farrell

Recanatesi

Lajoie

et al. 2019

Preprint

View full text Add to dashboard Cite

Recordings of neural circuits in the brain reveal extraordinary dynamical richness and high variability. At the same time, dimensionality reduction techniques generally uncover lowdimensional structures underlying these dynamics when tasks are performed. In general, it is still an open question what determines the dimensionality of activity in neural circuits, and what the functional role of this dimensionality in task learning is. In this work we probe these issues using a recurrent artificial neural network (RNN) model trained by stochastic gradient descent to discriminate inputs. The RNN family of models has recently shown promise in revealing principles behind brain function. Through simulations and mathematical analysis, we show how the dimensionality of RNN activity depends on the task parameters and evolves over time and over stages of learning. We find that common solutions produced by the network naturally compress dimensionality, while variability-inducing chaos can expand it. We show how chaotic networks balance these two factors to solve the discrimination task with high accuracy and good generalization properties. These findings shed light on mechanisms by which artificial neural networks solve tasks while forming compact representations that may generalize well.

show abstract

Section: Conclusion and Discussionmentioning

confidence: 99%

Dynamic compression and expansion in a classifying recurrent network

Farrell

Recanatesi

Lajoie

et al. 2019

Preprint

View full text Add to dashboard Cite

show abstract

“…To test the computational implications of trajectory divergence, we trained recurrent neural networks 630 with an atypical approach. Rather than training networks to produce an output, we trained them to 631 autonomously follow a target internal trajectory 38,51 . We then asked whether networks were able to 632 follow those trajectories from beginning to end, without the benefit of any inputs indicating when to…”

Section: Trajectory-constrained Neural Networkmentioning

confidence: 99%

Neural trajectories in the supplementary motor area and primary motor cortex exhibit distinct geometries, compatible with different classes of computation

Russo

Khajeh

Bittner

et al. 2019

Preprint

View full text Add to dashboard Cite

The supplementary motor area (SMA) is believed to contribute to higher-order aspects of motor control.To examine this contribution, we employed a novel cycling task and leveraged an emerging strategy: testing whether population trajectories possess properties necessary for a hypothesized class of computations. We found that, at the single-neuron level, SMA exhibited multiple response features absent in M1. We hypothesized that these diverse features might contribute, at the population level, to avoidance of 'population trajectory divergence' -ensuring that two trajectories never followed the same path before separating. Trajectory divergence was indeed avoided in SMA but not in M1. Network simulations confirmed that low trajectory divergence is necessary when guidance of future action depends upon internally tracking contextual factors. Furthermore, the empirical trajectory geometryhelical in SMA versus elliptical in M1 -was naturally reproduced by networks that did, versus did not, internally track context. 2Relative to primary motor cortex (M1), SMA activity is less coupled to actions of a specific body part 10-12 .3 Instead, SMA computations appear related to learned sensory-motor associations 11 , reward 4 anticipation 13 , internal initiation and guidance of movement 3,14 , movement timing 15,16 , and movement 5 sequencing 7,17,18 . Single-neuron responses in SMA reflect a variety of task-specific contingencies. For 6 example, in a sequence of three movements, a neuron may burst only when pulling precedes pushing. 7Another neuron might respond before the third movement regardless of the particular sequence 19 . 8 Different response features are observed in different tasks. Single SMA neurons exhibit a mixture of 9 ramping and rhythmic activity during an interval timing task 20 , and the SMA population exhibits 10 amplitude-modulated circular trajectories during rhythmic tapping 21 . A common thread linking prior 11 studies is that SMA computations are hypothesized to be critical when pending action depends upon 12 internal, abstract, and/or contextual factors. An important challenge is linking these high-level ideas to 13 network-level implementations. What general properties should activity exhibit in networks performing 14 the hypothesized class of computations? 15There exist many quantitative methods for relating population activity and computation (e.g., [22][23][24] ). 16These include decoding key hypothesized signals (e.g., via regression 25 ), or directly comparing empirical 17 and simulated population activity (e.g., via canonical correlation 26 ). An emerging strategy is to consider 18 the geometry of the population response: the arrangement of population states across conditions 23,27,28 19 and/or the shape traced by activity in neural state-space 15,26,[29][30][31][32][33][34][35][36] . A given geometry may be consistent 20 with some types of computation but not others 37 . An advantage of this approach is that it is sometimes 21 possible to measure geometric properties that are expected to hold f...

show abstract

“…Moreover, the trajectories are chaotic, and so inherently unstable and not robust to noise. Although, recent theoretical work 10,14 has demonstrated ways of making them robust.…”

Section: Introductionmentioning

confidence: 99%

Low dimensional dynamics for working memory and time encoding

Cueva

Saez

Marcos

et al. 2018

Preprint

Self Cite

View full text Add to dashboard Cite

Our decisions often depend on multiple sensory experiences separated by time delays. The brain can remember these experiences and, simultaneously, estimate the timing between events. To understand the mechanisms underlying working memory and time encoding we analyze neural activity recorded during delays in four experiments on non-human primates. To disambiguate potential mechanisms, we propose two analyses, namely, decoding the passage of time from neural data, and computing the cumulative dimensionality of the neural trajectory over time. Time can be decoded with high precision in tasks where timing information is relevant and with lower precision when irrelevant for performing the task. Neural trajectories are always observed to be low dimensional. These constraints rule out working memory models that rely on constant, sustained activity, and neural networks with high dimensional trajectories, like reservoir networks. Instead, recurrent networks trained with backpropagation capture the time encoding properties and the dimensionality observed in the data.

show abstract

full-FORCE: A target-based method for training recurrent networks

Cited by 130 publications

References 31 publications

Dynamic compression and expansion in a classifying recurrent network

Dynamic compression and expansion in a classifying recurrent network

Neural trajectories in the supplementary motor area and primary motor cortex exhibit distinct geometries, compatible with different classes of computation

Low dimensional dynamics for working memory and time encoding

Contact Info

Product

Resources

About