CSI 2022 Presentation Comment Sheet

	A	B	C	D	E	F
1	Draft Version	Slide Number	Commentor	Comments	Applied ?	Reply
2	1	3	Christophe	"Masters" --> "Master" (two times)
3	1	4	Christophe	The définition of NN seems counter intuitive : are they made up of layers of artificial neurons or of matrices and vectors? Starting with the figure would be better I think.		I added some slides for the explanation
4	1	all	Christophe	You miss points at the end of many sentences
5	1	6	Christophe	"your output" --> "the output"
6	1	7	Christophe	You have not expressed a problem, just facts; what is the problem?		These computations are expensive in terms of time and memory making them hard to use in real-world scenarios
7	1	8	Christophe	"aimed at working on NN acceleration" --> "aims at accelerating NNs"
8	1	8	Christophe	Learning and inference stages should have been defined before, when introducing NNs.		I added a point on slide 5
9	1	10	Christophe	Seems incomplete ...		True. Covered by point 1
10	1	13 & 16	Christophe	"will different" --> "with different" "with even" --> "even with"
11	1	15	Christophe	Maybe you could just explain what is dynamic instead of putting again all the items and figure from the static slide?		I do think the figure is important to remind them about how we are doing it. There is a bullet point describing what dynamic means
12	1	19	Christophe	"Using a code generated based on" --> "Using" "evolution" --> "evaluation" "polynomial evaluation" --> "polynomial approximation"		1. I think you mean "generator" since it was "generated"
13	1	19	Christophe	This slide implicitely implies you generated a single approximation for the lowest working precision, i.e., you can use it only for static strategies. Is that true?		For now that's true. We don't have dynamic switching with the code generated yet
14	1	25	Christophe	"Publish" --> "Complete experiments and consolidate theoretical explanations and then publish"
15	1	25	Christophe	"a goal" --> "the goal"
16	1	26	Christophe	Since there is a single item in each category, the titles could be singular
17	1	4-5	Silvui	"What are they" -> "What are they?"
18	1	4	Silvui	I don't like the definition here either and it would be better to start with an image that gives the broad idea of what a neural network is. After you give the image representation, you can talk about the general form of a layer (i.e. matrix multiplications and activation function evaluations)
19	1	all	Silvui	I'm missing a context for your talk. You start directly with neural networks and what they are. The question you should be asking at the beginning is what are you trying to do and why? Then, when you go into what you've done so far, that will consist of explaining how you do it.		Yes i thought about it. However during my talk i will layout the flow. I want to introduce NN because all my presentation has references to NN
20	1	6	Silvui	This slide is vague, you should differentiate between inference and training. And I would use an image to show how a DNN works.
21	1	7	Silvui	It would be great to have a more concrete idea of how the size of DNNs has been increasing in recent years; e.g. can you find a plot that shows how the size of DNNs has been increasing for various domains/applications? It's not really clear what is the problem you are trying to solve.
22	1	8	Silvui	LeanAI is focused on training acceleration, and not on inference
23	1	11	Silvui	"We ran three experiments" -> I would formulate this differently, you ran much more than three experiments ;-) You can say something like "We worked on three main scenarios/topics"		We worked on three main strategies
24	1	17	Silvui	You need a way to show on the plots when you do the precision switch, similar to what you have on slide 18
25	1	9	Wassim	Add images to make it clearer ( neural network and
26	1	9	Wassim	introduce different acitvation functions with graphs
27	1	13	Wassim	Introduce MpTorch
28	1	all	Wassim	Make english more formal
29	1	14	Wassim	Add remark about forward & backward being different quantization
30	1	20	Wassim	Illustrate the code generator
31	1	21	Wassim	The code generator should not be in the box.
32	1	26	Wassim	Make the Plan as time stage
33	1	18	Wassim	In the test graph "e" is small in "Epoch"
34	1	all	Christophe	Increase font for slide numbers		I will leave this until the end because I have to do them manually
35	1	14 & 17 & 18	Christophe	You should say a word about timing there: the aim is "accelerating NNs", is it achieved?		These are the software simulated experiments. We don't get any time benefits. Do you think I should state that ?
36	1	16	Christophe	Anastasia's and Silviu's opinions needed here : Here you will trigger a question: why did you add a Resnet for this experiment? Since the explanation is purely chronological (if I remember correctly), you should perhaps use it also for static and display figures for it too?
37	1	17 & 18	Christophe	Since dynamic precision requires monitoring, what is the overhead wrt static precision? You should say a word about that.		Monitor the training and testing performance by checking if the accuracy is stalled after a certain number of epochs ( negligible overhead )
38	1	21	Christophe	Why only Resnet?		After dynamic we only did the experiments on the ResNet
39	1	22	Christophe	You will have to explain why the low precision needs less epochs, and if the time gain is only due to this.		Good point. Well this is due to the EarlyStopping mechanism. I will add that to the graph as an annotation
40	1	28	Christophe	If would recommend including these references in footnote in the slides where they are cited. Conversely, you could list here the most important items from the bibliography you have covered in the past months...
41	1	10	Silvui	This slide is definitely incomplete. When putting references in a slide, I would add them as a footnote to the page. You need to be more complete in your presentation of the state of the art, and in general, you should try to constantly follow if there are new works appearing that use low/mixed-precision computing for training acceleration.		OK
42	1	12	Silvui	You should be clearer about the fact that you are quantizing to a lower precision; you should have a slide in the beginning about quantization, since lowering the numerical precision is the main way you want to accelerate training in your work		Not sure if still applicable
43	1	13 & 16	Silvui	You should do more experiments with more networks and datasets. Why are those theoretical results and not experimental results? What is the theory behind them? You talk about quantizing activation functions, but you never say which activation functions are you quantizing in your experiments (e.g. GeLU, tanh, sigmoid?). You should also run the static and dynamic quantization experiments on the same networks and datasets, since this would allow you to compare the two approaches. You also need a slide where you compare the two approaches (static vs dynamic) and you need to discuss the possible advantages and disadvantages of both.
44	1		Wassim	Add comparison slide between static and dynamic quantization
45	1	19	Wassim	Add a second experiment for ResNet
46	2	15	Wassim	Add Literature review
47	2	25	Wassim	Add annotation on the test accuracy graph
48	2	29	Wassim	Add better graphs
49	2	5	Wassim	Add applications of DL in the space below
50	2	6	Wassim	increase images slide
51	2	7	Wassim	Remove the bracket and the equation below
52	2	7	Wassim	Add input to layers notation
53	2	8	Wassim	Put the formula under the big box instead of under the layer naem
54	2	8	Wassim	a l -1 remove paranthesis
55	2	8	Wassim	Specify that there are multiple types of layers
56	2	8	Wassim	Say that W & b are weights and biases
57	2	8	Wassim	Decouple the layers and activation functions
58	2	9	Wassim	In the derivative, put the weight update
59	2	9	Wassim	Rotate the images and transition from forward & backward instead of 2 images
60	2	10	Wassim	Redundant, add it to slide 9
61	2	11-12	Wassim	Merge them,		To be discussed
62	2	12	Wassim	Fix wording
63	2	12	Wassim	Add leanAI partners
64	2	15	Wassim	How do you position yourself in the state of the art
65	2	16	Wassim	After literature review add idea of what you are doing
66	2	all	Wassim	Move from quantization to uniform precision switch
67	2	all	Wassim	Remove software simulation
68	2	19	Wassim	Move image from 20 to 19
69	2	21	Wassim	Can be removed and just said on slide 20
70	2	23	Wassim	Change text
71	2	25	Wassim	Remove arrows
72	2	28	Wassim	Low Accuracy instead of lwo precision
73	2	All	Wassim	Dynamic Quantization -> Dynamic Precision Switching
74	2	27	Wassim	All some text about the generate code
75	2	31	Wassim	Remove the summary		I still feel like that it's good to have it
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100