About Me

As a big AI enthusiast, I can't wait to see the next 'big thing' in AI. I strongly believe that for any serious progress towards next-generation AI, we need to fixate less on backprop and instead try more to understand and reverse-engineer biological neural circuits. If you think this is an insane challenge, I agree. But it's worth the effort. To help make this a reality, I joined the Neural Coding and Brain Computing Unit at OIST, where I am now investigating how the brain is able to learn so quickly and efficiently.

My recent research has been under the following themes:

  • Self-organized cell assemblies as a substrate of episodic memory: Model of asscociative memory under a biologically-inspired STP-dependent symmetric STDP learning rule.
  • Preplay as a learning booster: I am modeling how new knowlede can be bootstrapped from preexisting topological and temporal structure of a contineously plastic binary neural network.
  • Role of spontaneous neural activity in memory: Spontaneous neural activity in animals has been shown to reflect sensory priors. I aim to develop a biologically plausible simplified SNN model that can reliably store sensory priors and use them for task performance.

Before I joined the NCBC, I had spent four months at the Cognitive Neurorobotics Unit where I 'taught' a humanoid robot to perform a reach-and-grasp task by combining a limited set of learned motor primitives into diverse and novel motion trajectories. This work was inspired by the stochastic PV-RNN architecture.


Projects

Emergence and Functional Role of Spontaneous Activity in Spiking Neural Networks. Bilogical neural networks maintain a certain level of spontaneous activity (SA) at all times, even in the absence of external stimuli. If SA were nothing more than just noise due to random discharges of neurons, as originally thought by some, it would be a big waste of energy given that mammalian brains consume a considerable amount of energy even in quiescent states like sleep. Recent evidence suggests that SA is implicated in memory, spatial navigation and learning, but precisely what its computaional role is remains unclear. In this project I investigate, using biologically plausible and experimentally constrained spiking recurrent network models, how SA emerges and persists as well as how it can enable and/or enhance memory function. The ultimate goal of this project is to explore if and how structured SA can be useful in next-generation AI.

Neurobarometer was a project funded by Neurontrend LLC and the Russian Venture Company to develop an algorithm for objective estimation of people's affective state for neuromarketing research. Based our method (now patented), we built an easy-to-use application that provides a bias-free estimate of respondents' opinion of a product sample based on real-time 32-channel EEG. The first prototype was built in 2019, and the application is now ready and sold to enterprise customers. My responsibilities in this project included developing optimal data preprocessing pipelines and a robust prediction model. [YouTube]

Human-Robot Interaction: This project explored the ability of stochastic recurrent neural networks trained on a small set of motor primitives to generate meaningful novel patterns with limited corrective feedback from the experimenter. [YouTube]


Patents

RF Patent 2747571. Electroencephalographic method and system of objective estimation of listeners’ reaction to audio content based on a range of voluntary affective categories. [Full text]



Publications

Peer-reviewed Journal Publications

Testing the Efforts Model of Simultaneous Interpreting
Roman Koshkin, Yury Shtyrov, Andriy Myachykov & Alex Ossadtchi
PLoS ONE 13(10): e0206129
PDF Abstract Press Bibtex Github
We utilized the event-related potential (ERP) technique to study neural activity associated with different levels of working memory (WM) load during simultaneous interpretation (SI) of continuous prose. The amplitude of N1 and P1 components elicited by task-irrelevant tone probes was significantly modulated as a function of WM load but not the direction of interpretation. Furthermore, the latency of the P1 increased significantly with WM load. The WM load effect on N1 latency, however, did not reach significance. Larger negativity under lower WM loads suggests that more attention is available to process the source message, providing the first electrophysiological evidence in support of the Efforts Model of SI. Relationships between the direction of interpretation and median WM load are also discussed.
Commentary: Functional Connectivity in the Left Dorsal Stream Facilitates Simultaneous Language Translation: An EEG Study
Roman Koshkin & Alex Ossadtchi
Frontiers in Human Neuroscience, 11(2), 273
PDF Bibtex

Conference Proceedings

Working Memory Load in Simultaneous Language Interpretation: An ERP Study.
Roman Koshkin, Alex Ossadtchi
4th Conference "Cognitive Science in Moscow: New Research.
PDF Abstract
We utilized the event-related potential (ERP) technique to study neural activity associated with different levels of working memory (WM) load during simultaneous language interpretation (SLI). We pioneered the use of the technique on conference interpreters articulating overtly. The amplitude of the N1 component elicited by task-irrelevant tone probes was significantly modulated as a function of WM load but not the direction of interpretation. The N1 amplitude decreased with load, suggesting shallower processing under high WM load regardless of the direction.

Presentations and Talks

Testing One Aspect of the Efforts Model of Simultaneous Interpreting: An ERP Study.
Roman Koshkin, Yury Shtyrov, & Alex Ossadtchi.
Workshop "Neurobiology Of Speech And Language", October 27-29, 2017, Saint-Petersburg, Russia
Poster Abstract
Due to the inherent complexity of simultaneous interpreting (SI), testing its theoretical models in a simple, but ecologically valid experimental paradigm has been problematic. We attempted to overcome some of the associated challenges using a novel method for WM load estimation in conjunction with the event-related potentials (ERP) technique. Specifically, we tested the prediction of the Efforts Model of SI (Gile, 1988) that increased WM load impairs the processing of the source message during SI.Consistent with the model, the N1 amplitude that we used as an index of attention was significantly modulated as a function of WM load. Negativity in the N1 range decreased at higher levels of WM load, suggesting shallower processing of the source message under high WM load. Our findings represent the first electrophysiological evidence in support of the Efforts Model.
N1 ERP As an Index of Depth of Processing In Simultaneous Interpreting.
Roman Koshkin, Alex Ossadtchi & Yury Shtyrov.
Communication, Computation, and Cognitive Processes, September 28-29, 2016, Moscow, Russia
PDF Abstract
Researchers have extensively studied simultaneous interpretation (SI), but understanding how the brain manages the limited resources of attention and working memory (WM) systems under such extreme conditions of language control has remained elusive. Here, we aim to use the event-related potentials (ERP) to investigate the interplay of attention and WM load during simultaneous interpretation of real speech in an ecologically valid overt production paradigm. Previous research using simple dichotic listening paradigms showed the N1 ERP component to be modulated by attention thus making it a suitable temporally precise index of attention. Specifically, we test the hypothesis that at larger WM loads, attention to the source message is markedly reduced as it gets redeployed towards processing the backlog of previous information. If this is borne out, translation fidelity is more likely to be compromised during periods of higher WM load. Additionally, we examine if concurrent articulation degrades attention to, and processing of, the source message in SI. Although such a setup requires non-trivial artifact correction techniques pushing the limits of the ERP method, the pattern of results based on one participant’s data suggests an effect of WM load on the N1 amplitude , which is in agreement with our original hypothesis.
Working Memory Load In Simultaneous Language Interpretation: An ERP Study.
Roman Koshkin, Alex Ossadtchi & Yury Shtyrov.
IEEE International Symposium «Video and Audio Signal Processing in the Context of Neurotechnologies», June 26-30, 2017, Saint Petersburg, Russia
PDF Abstract
We utilized the event-related potential (ERP) technique to study neural activity associated with different levels of working memory (WM) load during simultaneous language interpretation (SLI). We pioneered the use of the technique on conference interpreters articulating overtly. The amplitude of the N1 component elicited by task-irrelevant tone probes was significantly modulated as a function of WM load but not the direction of interpretation. The N1 amplitude decreased with WM load suggesting shallower processing under high WM load regardless of the direction. Using our novel projection-based method we identified otherwise hidden WM load-dependent regularities in the P3 range. The results are discussed in terms of the Efforts Model of simultaneous language interpreting.
Localizing Hidden Regularities With Known Temporal Structure in the EEG Evoked Response Data.
Alexandra Kuznetsova, Roman Koshkin, & Alex Ossadtchi.
IEEE International Symposium «Video and Audio Signal Processing in the Context of Neurotechnologies», June 26-30, 2017, Saint Petersburg, Russia
PDF Abstract
We describe a novel data driven spatial filtering technique that can be applied to the evoked potentials in the EEG data in order to find statistically significant hidden differential activations, which cannot be found by standard single-channel analysis. The underlying optimization problem is formulated as a generalized Rayleigh quotient maximization problem. The technique is based on the known morphological characteristics of the response: the optimal filter maximizes the difference in the target interval when the component typically occurs and at the same time minimizes the difference in the flanker interval. The technique is equipped with a relevant randomization-based statistical test to assess the significance of the discovered phenomenon. The performance of the proposed method was evaluated with the simulated ERP data, the results are compared with the competing ICA-based method. Furthermore, we describe an application of the proposed method to the EEG data acquired in two studies: study devoted to the simultaneous language interpreting (group analysis) and analysis of the auditory neuroplasticity (single subject application). We show how the differential components can be detected after filtration and support our results with the permutation statistical test, topographies analysis and single-trial evidence.
Does High WM load Disrupt Listening in Simultaneous Interpreting?
Roman Koshkin Alex Ossadtchi & Yury Shtyrov
Higher School of Economics, April 27, 2017.
Slides

Miscellaneous

I was a master's student of Dr. Alex Ossadtchi at at the Center for Bioelectric Interfaces.
As a hobby, I like playing with time-series prediction for stock and crypto trading as well as building my own bots for high-frequency trading (If you are a pro trader, don't worry -- none of my bots make a lot of money. Yet.). I also enjoy making useless (but irresistibly interesting) things with generative adversarial networks (GANs).