CNS*2020 Online has ended
Welcome to the Sched instance for CNS*2020 Online! Please read the instruction document on detailed information on CNS*2020.
Back To Schedule
Monday, July 20 • 9:00pm - 10:00pm
P84: Brainpower – Investigating the role of power constraints in neural network plasticity

Log in to save this to your schedule, view media, leave feedback and see who's attending!

Feedback form is now closed.
Poster meeting: meet.google.com/zvi-wofn-foh

Silviu Ungureanu
, Mark van Rossum

The central nervous system consumes approximately 20W of metabolic power in humans[1]. This is used for neural communication, and also for neural plasticity and the formation of new memories. Persistent forms of plasticity in particular consume so much energy that under sudden food scarcity, associative learning significantly reduces lifespan of fruit flies[2].

It is reasonable therefore that neural plasticity has evolved to learn at minimal power. However, how this changes plasticity and learning is not known. While previous work has considered an energy constraint[6], a power constraint might be more biological as, unlike many other tissues, the brain cannot store energy. A power constraint might be able to explain why plasticity induction requires a refractory time before it can be induced again[3], as well as spatial competition in plasticity between synapses and neurons[4][5].

Here, we developed a computational model of plasticity to examine the effect of a power constraint on plasticity dynamics. We first use a standard perceptron augmented with two types of synaptic weights: an inexpensive transient, decaying component and a costly long-term component, formed by the simultaneous consolidation of all the transient weights. We further assume that the brain attempts to consolidate new memories as soon as it is able to. Hence, the interval between consolidation events is limited and synaptic consolidation events occur at a fixed frequency, representing the refractory period caused by a dearth of energy. Higher consolidation frequencies correspond to more available power, and vice-versa. The perceptron is trained on a random-generated set of binary patterns until it correctly learns the output value for each pattern.

Results show that the power in the system has a significant impact on the training time. Unexpectedly, increasing the period between consolidations - thus reducing power - can reduce the required number of epochs by as much as 30%, depending on the strength of the weight decay, the number of patterns P in the training set, and the number of synapses N. Further increasing the period between consolidations increases the training time. This increase occurs not gradually, but in a staircase pattern, peaking whenever the period is 0 modulo P, Fig.1.

The consequences of a power constraint are further explored in a multi-layer neural network and extended to a probabilistic model where the probability of consolidation increases proportional to the time since the previous consolidation event.

In summary, our results show that incorporation of a metabolic power constraint in synaptic plasticity can lead to important changes in the learning dynamics.


1. Attwell D, Laughlin SB. An energy budget for signaling in the grey matter of the brain. J Cerebral Blood Flow 2001, 21, 1133. 2. Mery F, Kawecki TJ. A cost of long-term memory in Drosophila. Science. 2005, 308, 1148. 3. Kramár EA, Babayan AH, Gavin CF, et al. Synaptic evidence for the efficacy of spaced learning. PNAS 2012, 109, 5121. 4. Sajikumar S, Morris RG, Korte M. Competition between recently potentiated synaptic inputs reveals a winner-take-all phase of synaptic tagging and capture. PNAS 2014, 111,12217. 5. Josselyn SA, Tonegawa S. Memory engrams: Recalling the past and imagining the future. Science. 2020, 367, 6473. 6. Li, H.L., Van Rossum, M.C. Energy efficient synaptic plasticity. Elife. 2020, e50804.


Silviu Ungureanu

PhD Student, School of Psychology, University of Nottingham

Monday July 20, 2020 9:00pm - 10:00pm CEST
Slot 02