Classes & TalksTomorrow
Dissertation Proposal: Ikteder Akhand Udoy
Tuesday, November 17 · 2:45-4:45 p.m. · Grant Avenue Annex (GAA)
- When
- Tuesday, November 17 · 2:45-4:45 p.m.
- Where
- Grant Avenue Annex (GAA)1015 South Grant Avenue · Boise, ID · 83706
About this event
Title: Toward Multiplier-Free Linear Layers and Memory-Efficient Local Training
Program: Computing Ph.D.
Committee Chair: Omiya Hassan
Committee: Omiya Hassan, Kurtis Cantley, Jun Zhuang
Abstract:
The computational and memory demands of neural networks limit their deployment and
adaptation on resource-constrained hardware. This proposal investigates the combination of
SPARQ, a power-of-two quantization-aware training method, and Hidden-Teacher Forward
(HTF), a group-local learning method inspired by the Forward-Forward algorithm. Published
results from an earlier two-term SPARQ configuration, reported under the name Po2-QAT,
demonstrate that selected language-model weights can be trained toward shift-friendly
representations, with a measured quality cost. Preliminary results from a submitted HTF
manuscript demonstrate end-to-end training while revealing tradeoffs in quality, memory use,
and runtime. The proposed research will extend SPARQ to a one-term-or-zero codebook and
evaluate its interaction with HTF through matched language-model experiments. It will then
investigate reduced-state adaptation on a microcontroller and shift/add inference on an FPGA.
Model quality, training memory, runtime, device latency, and energy will be measured against
relevant baselines. The study aims to establish when the combined method improves the
quality-resource tradeoff and when its constraints or overhead outweigh the hardware benefits.
Listed by Boise State University (Localist), gathered by the Caldwell events calendar.
Does something look wrong?
Location
Grant Avenue Annex (GAA)1015 South Grant Avenue · Boise, ID 83706
Get directions