appreciate why we need efficiency in deep learning models today, how to think about it in terms of metrics that you care about, and finally the tools at your disposal to achieve what you want. The subsequent accompanying the given prompts. Both these models have been deployed in production. BERT is used in Google Search to improve relevance of results, and GPT-3 is available as an API for interested users to consume Using the sensitive tweet classifier example, during the deployment phase the user will be concerned about the inference efficiency and should be aware of what is the inference latency per tweet, peak RAM

of humans but of mind-numbing behavior." - Stewart Butterfield, Founder (Slack) We have talked about a variety of techniques in the last few chapters to improve efficiency and boost the quality of deep plethora of choices that we face when training a deep learning model in the computer vision domain. A Search Space for n parameters is a n-dimensional region such that a point in such a region is a set of each of those parameters. The parameters can take discrete or continuous values. It is called a "search" space because we are searching for a point in which minimizes (or maximizes) an Evaluation Function

"system", "content": "You are a helpful assistant."}, {"role": "user", "content": "Tell me something about large language models."} ], }' 或者您可以按照下面所示的方式，使用 openai Python 包中的 Python 客户端： from openai import "system", "content": "You are a helpful assistant."}, {"role": "user", "content": "Tell me something about large language models."}, ] ) print("Chat response:", chat_response) 1.2.3 下一步 现在，您可以尽情探索 Qwen "system", "content": "You are a helpful assistant."}, {"role": "user", "content": "Tell me something about large language models."} ], }' 或者你可以按照下面所示的方式，使用 openai Python 包中的 Python 客户端： from openai import

edu.cn September 6, 2023 Feng Li (SDU) Overview September 6, 2023 1 / 57 Lecture 1: Overview 1 About the Course 2 Machine Learning: What and Why? 3 Categories of Machine Learning 4 Some Basic Concepts training examples selected by a “benevolent” teacher. “Near miss” examples Learner can query an oracle about class of an unlabeled example in the environment Learner can construct an arbitrary example and query guidance. Feng Li (SDU) Overview September 6, 2023 15 / 57 Applications of Machine Learning Document Search Given counts of words in a document, determine what its topic is. Group documents by topic without

or ensemble of models to smaller models. The obvious question at this point is: why are we talking about them in the same breadth as efficiency? To answer this question, let’s break down the two prominent demonstrates sample efficiency between two model training setups. The sample efficient model achieves about the same accuracy, but reaches that point in fewer epochs, hence needing fewer samples. Distillation Now, there can be a few different options available to us, based on what we want: 1. We only care about reaching the accuracy goal of 80%: In this case, it is perfectly fine to take the lower labeling and

limited number of labeled examples for fine-tuning since the model already knows the general concepts about language, and use the same model across many tasks. Model reuse by itself also is a powerful attribute common self-supervised learning into two broad steps: 1. Pre-training: This step teaches the model about the world it is operating in (language, vision, multimodal) through certain tasks which ensure that in Kaggle and Google Colab (apart from the paid service on Google Cloud). We will be talking more about TPUs in Chapter 10. For now, you can follow our lead. You can also adapt the code to run on GPUs if

你 好, 梯 度 主讲人：龙良曲 Gradient Descent ▪ ???? = ?2 ∗ sin(?) How about Linear Equations ▪ ? = ? ∗ ? + ? ▪ 1.567 = w * 1 + b ▪ 3.043 = w * 2 + b ▪ W = 1.477 ▪ B = 0.089 Closed Form Solution With ? How to optimize ▪ ???? = σ? ? ∗ ?? + ? − ?? 2 ▪ Minimize ???? ▪ ?′ ∗ ? + ?′ → ? Heuristic Search Convex Optimization https://spin.atomicobject.com/2014/06/24/gradient-descent-linear-regression/

params = {‘kernel’:[‘linear’, ‘rbf’], ‘C’:[1, 10]} grid_search = GridSearchCV(svc, params, cv=5) grid_search.fit(X_train, y_train) grid_search.best_params_ 在参数网格上进行穷举搜索，方法简单但是搜索速度慢（超参数较多时），且不 容易找到参数空间中的局部最优 {‘kernel’:[‘linear’, ‘rbf’], ‘C’:randint(1, 20)} random_search = RandomizedSearchCV(svc, param_dist, n_iter=10) random_search.fit(X_train, y_train) random_search.best_params_ 在参数子空间中进行随机搜索，选取空间中的100个点进行建模（可从

analysis, network/tribe analysis Netflix • Recommendation engine Pinterest • Image recognition search Fraud.net • Detect online payment fraud DataXu • Leverage automated & unattended ML at large scale compile on … Amalgamation Runs in browser with Javascript The first image for search “dog” at images.google.com Outputs “beagle” with prob = 73% within 1 sec Deep RL | Playing

PS的参数放置策略 • Ps分布式分片的均衡，避免分片大小不一致  NN网络矩阵按行切分，解决请求包不均衡问题  特征按照Hash方式分布式存储 • 模型并行调超参  grid search  random search PS的多模型训练 • 提高内存使用效率  model group内共享特征key的存储 • 超大规模模型 -> 高扇出的分布式PS • 长尾效应：单个分片的抖动（网络、CPU）对请求影响变大