An official website of the United States government
We consider the post-training quantization problem, which discretizes the weights of pre-trained deep neural networks without re-training the model. We propose multipoint quantization, a quantization method that approximates a full-precision weight vector using a linear combination of multiple vectors of low-bit numbers; this is in contrast to typical quantization methods that approximate each weight using a single low precision number. Computationally, we construct the multipoint quantization with an efficient greedy selection procedure, and adaptively decides the number of low precision points on each quantized weight vector based on the error of its output. This allows us to achieve higher precision levels for important weights that greatly influence the outputs, yielding an 'effect of mixed precision' but without physical mixed precision implementations (which requires specialized hardware accelerators). Empirically, our method can be implemented by common operands, bringing almost no memory and computation overhead. We show that our method outperforms a range of state-of-the-art methods on ImageNet classification and it can be generalized to more challenging tasks like PASCAL VOC object detection.
more »
« less
Predicting Performance and Accuracy of Mixed-Precision Programs for Precision Tuning
- Award ID(s):
- 1750983
- PAR ID:
- 10541022
- Publisher / Repository:
- ACM
- Date Published:
- ISBN:
- 9798400702174
- Page Range / eLocation ID:
- 1 to 13
- Format(s):
- Medium: X
- Location:
- Lisbon Portugal
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Stable precision grips using the fingertips are a cornerstone of human hand dexterity. However, our fingers become unstable sometimes and snap into a hyperextended posture. This is because multilink mechanisms like our fingers can buckle under tip forces. Suppressing this instability is crucial for hand dexterity, but how the neuromuscular system does so is unknown. Here we show that people rely on the stiffness from muscle contraction for finger stability. We measured buckling time constants of 50 ms or less during maximal force application with the index finger—quicker than feedback latencies—which suggests that muscle-induced stiffness may underlie stability. However, a biomechanical model of the finger predicts that muscle-induced stiffness cannot stabilize at maximal force unless we add springs to stiffen the joints or people reduce their force to enable cocontraction. We tested this prediction in 38 volunteers. Upon adding stiffness, maximal force increased by 34 ± 3%, and muscle electromyography readings were 21 ± 3% higher for the finger flexors (mean ± SE). Muscle recordings and mathematical modeling show that adding stiffness offloads the demand for muscle cocontraction, thus freeing up muscle capacity for fingertip force. Hence, people refrain from applying truly maximal force unless an external stabilizing stiffness allows their muscles to apply higher force without losing stability. But more stiffness is not always better. Stiff fingers would affect the ability to adapt passively to complex object geometries and precisely regulate force. Thus, our results show how hand function arises from neurally tuned muscle stiffness that balances finger stability with compliance.more » « less
-
null (Ed.)Antibodies, particularly of the immunoglobulin G (IgG) isotype, are a group of biomolecules that are extensively used as affinity reagents for many applications in research, disease diagnostics, and therapy. Most of these applications require antibodies to be modified with specific functional moieties, including fluorophores, drugs, and proteins. Thus, a variety of methodologies have been developed for the covalent labeling of antibodies. The most common methods stably attach functional molecules to lysine or cysteine residues, which unavoidably results in heterogeneous products that cannot be further purified. In an effort to prepare homogeneous antibody conjugates, bioorthogonal handles have been site-specifically introduced via enzymatic treatment, genetic code expansion, or genetically encoded tagging, followed by functionalization using bioorthogonal conjugation reactions. The resulting homogeneous products have proven superior to their heterogeneous counterparts for both in vitro and in vivo usage. Nevertheless, additional chemical treatment or protein engineering of antibodies is required for incorporation of the bioorthogonal handles, processes that often affect antibody folding, stability, and/or production yield and cost. Accordingly, concurrent with advances in the fields of bioorthogonal chemistry and protein engineering, there is growing interest in site-specifically labeling native (nonengineered) antibodies without chemical or enzymatic treatments. In this review, we highlight recent strategies for producing site-specific native antibody conjugates and provide a comprehensive summary of the merits and disadvantages of these strategies.more » « less
-
null (Ed.)We discuss the improvements that the ILC can make in precision electroweak observables based on studies with the ILD detector concept. These include observables from WW production at a centre of mass energy of 250 GeV and above, and especially from a dedicated stage of running at the Z pole. These improvements take advantage of the ILC capabilities for polarized electron and positron beams, and an accelerator design that accommodates data-taking at a wide range of beam energies. The studies include experimental considerations evaluated in the context of the ILD detector concept and discussion of experimental strategies targeted at controlling especially systematic uncertainties associated with the center-of-mass energy.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Conference Paper:
- https://doi.org/10.1145/3597503.3623338
