# #COSYNE18

## Tutorial

This year, we had the first tutorial session for COSYNE. From the beginning of COSYNE, there was a demand & plan for a tutorial session (according to Alex Pouget). There were 293 registered for the tutorial, and it was their first COSYNE for [155, 208] (95% CI) participants. Everybody who gave me feedback was very happy with Jonathan Pillow‘s 3.5 hour lecture (slides & code) on statistical modeling techniques for neural signals, so we are planning to run another tutorial next year at **#COSYNE19 Lisbon, Portugal**.

## Main meeting

**Basic stats**: 857 registrations, 709 abstracts submitted, 396 accepted (55.8%) which was increased from 330 at Salt lake city thanks to a bigger venue at Denver (curiously, it used to be where NIPS was up till 2000).

My biased **keywords** of #COSYNE18: neural manifold, state space, ring attractor, recurrent neural network. Those dynamical system views are going strong.

Below are just some notes for my future self.

**Tiago Branco, Computation of instinctive escape decisions (invited talk)**

Beautiful dissecting of escape decision and response circuit in a looming dark circle task. He showed that mSC (superior colliculus) is causally involved in threat evidence representation, and the immediate downstream dPAG (periaqueductal gray) causally represents escape decision Interestingly, the synaptic connection from mSC to dPAG is weak and unreliable (fig from bioRxiv), possibly contributing to the threshold computation for escape decision.

**Iain D. Couzin, Collective sensing and decision-making in animal groups: From fish schools to primate societies (Gatsby lecture)**

Ian showed how he went from theoretical models of swarm behavior to virtual reality for studying interacting fish (I thought this is how you spend money in science!), how the group can solve spatial optimization problems without any individual having access to the gradient (PNAS 2009), how collective consensus can go from following strongly biased minority to a “democratized” decision making as the swarm size increased (Science 2017), and also how the collective transitions from averaging to winner take all (fig from Trends CogSci 2009).

**I-80. Michael Okun, **Kenneth Harris**. Frequency domain structure of intrinsic infraslow dynamics of cortical microcircuits**

In the time scale of tens of seconds (infraslow), they showed that inter-spike intervals alone does not, but with matching power spectral density does explain much of the slow variations. (Spikes with matching spectra and ISI were generated using a variation of amplitude adjusted Fourier transform).

**I-15.** **Rudina Morina**, Benjamin Cowley, Akash Umakantha, Adam Snyder, Matthew Smith, Byron Yu. **The relationship between pairwise correlations and dimensionality reduction**

From paired recordings, the spike count correlation distribution is often reported as evidence of low-dimensional activity. From population recordings, factor analysis is often used as measures of neural dimensionality. How do these two relate? Both quantities only depend on the covariance matrix, hence they investigate how the mean and standard deviation of spike count correlation (rSC) relate to dimensionality as a function of shared variance using the generative model of factor analysis. They found an interesting trend: low-dim can be either large mean rSC with small std rSC or small mean and large std (which can be shown by rotating the loadings matrix).

**I-13. Lee Susman**, Naama Brenner, Omri Barak. **Stable memory with unstable synapses
**Using anti-Hebbian learning rule , they stored memories in limit cycles instead of stable fixed points as in traditional Hopfield networks. Also, they added a Chaotic non-stationary dynamics in the symmetric portion of the network which made the network continuously fluctuate and escape limit cycles (without losing the memory).

**T-14. Emily Mackevicius**, Andrew Bahle, Alex Williams, Shijie Gu, Natalia Denissenko, Mark Goldman, Michale Fee**. Unsupervised discovery of neural sequences in large-scale recordings
**Conventional low-rank decomposition of temporal data matrix cannot find sequential structure. They have extended the convolutional non-negative matrix factorization (Non-negative Matrix Factor Deconvolution; Smaragdis 2004, 2007) for neural data, and called it

**SeqNMF**. (MATLAB code on github; thanks for the quick bug fix @ItsNeuronal; bioRxiv 273128). It’s performance on syllable extraction and spike sequence detection was very nice.

**Timothy Behrens. Building models of the world for behavioural control (invited talk)**

I usually ignore fMRI talks, but the 6-fold symmetry of conceptual space was very cool. In the “canonical” bird neck & leg length space, OFC and other areas showed grid-network like signal modulations (Science 2016).

CATNIP lab had several posters on the 2nd day:

**II-2. David Hocker, Il Memming Park. Myopic control: A new control objective for neural population dynamics****II-7. Yuan Zhao, Il Memming Park. Accessing neural states in real time: recursive variational Bayesian dual estimation****II-62. Kathleen Esfahany, Isabel Siergiej, Yuan Zhao, Il Memming Park. Organization of neural population code in mouse visual system**

**Marlene Cohen. Understanding the relationship between neural variability and behavior (invited)
**If correlated variability is important, it should (1) be related to performance, (2) related to individual perceptual decisions, (3) be selectively communicated between brain areas. She showed that the first principal component of noise correlation, but not the signal encoding direction, is most correlated with the choice. This was just recently published in Science 2018.

**T-24. Caroline Haimerl, **Eero Simoncelli**. Shared stochastic modulation can facilitate biologically plausible decoding**

Noise correlation tends to be in the direction of strongest decoding signal for unknown reason (Lin et al. 2015; Rabinowitz et al. 2016). They used the neural response weighted by the shared gain modulation for decoding, which was near optimal.

**Byron Yu. Brain-computer interfaces for basic science (invited)
**They used BCI to study how the monkey can change its output in a short time scale within the “neural manifold”. Surprisingly, the neural repertoire (distribution of possible population firing patterns) does not shift nor change in shape, but mostly reassigns meaning! (Nature Neuroscience 2018) The animal can learn out of manifold perturbation as well, but that takes days (as detailed in

**Emily Oby**‘s talk (T-25) followed right after).

**T-26. Evan Remington, **Devika Narain, Eghbal Hosseini, Mehrdad Jazayeri. **Control of sensorimotor dynamics through adjustment of inputs and initial condition**

In a ready-set-go time interval production task with variable gain (animal has to reproduce 1.5 times the duration sometimes), the mean neural activity of the population forms a #neuralManifold. On the interval subspace, the two temporal gains produced identical mean trajectories, while on the gain subspace, they were separated. (bioRxiv 261214)

**Máté Lengyel. Sampling: coding, dynamics, and computation in the cortex (invited)**

If the population neural activity represents samples from the posterior, what neural dynamics would produce them (Rubin et al. 2015)? He showed that a stochastic stabilized supralinear network (SSN) with a ring architecture (not ring attractor) can sample and also reproduce neurophysiological temporal dynamics such as on/off-set response and quenching of variability (bioRxiv 2016). Also he trained an RNN to amortize inference of a simple Gaussian scale mixture model of vision, and the solution found by the RNN turns out to be non-detailed balance solution to sampling (as demonstrated by the anti-symmetric part of cross-correlation over time).

**T-37. Lucas Pinto**, David Tank, Carlos Brody, Stephan Thiberge. **Widespread cortical involvement in evidence-based navigation
**Using a transparent skull animal on Poisson towers VR decision-making task (Front. Behav. Neurosci. 2018), they found that many areas in the dorsal cortex (V1, SS, M1, RSC, mM2, aM2, etc) were correlated with accumulation of evidence. Further optogenetic inactivation of each area disrupted animal’s performance.

**Vivek Jayaraman. Navigational attractor dynamics in the Drosophila brain: Going from models to mechanism (invited)**

Beautiful work on the ellipsoid body–protocerebral bridge circuit and their computation involving bump attractor dynamics and path integration.

**Joni Wallis. Dynamics of prefrontal computations during decision-making (invited)**

Theta-oscillation phase of OFC locks to the trials when the reward criteria were linearly changing. A closed-loop microstimulation of OFC at the peak of theta disrupts learning, possibly due to disrupted theta-locked communication with hippocampus.

**III-108. Rainer Engelken, **Fred Wolf.** A spatiotemporally-resolved view of cellular contributions to network chaos**

They implemented a event-based recurrent spiking neural network that is so efficient that they can simulate a very large number (15 million) of neurons and study their dynamics. They quantified Lyapunov exponents efficiently and computed cross-correlation against the participation index.

**III-75. KiJung Yoon**, Xaq Pitkow.** Learning nonlinearities for identifying regular structure in the brain’s inference algorithm**

Loopy belief propagation often produces poor inference on non-tree graphical models. Can we do better by training a recurrent neural network to do amortized inference on graphs? The answer is yes, and it can generalize to larger networks and unseen graph structures.

**III-121. Dongsung Huh**, Terrence Sejnowski.** Gradient descent for spiking neural networks**

They derived a differentiable synapse which gradually responses to membrane voltage near threshold. The presynaptic neuron still spikes, but the differentiable synapse allows gradient descent training of the recurrent spiking neural network. During training they can slowly make the synapse tighter and tighter to finally reach a non-differentiable synapse (arXiv 2017).