Skript Flashcards
How do we define “artificial intelligence”?- Def.
Intelligence exhibited by non-biological systems
Artificial General Intelligence (AGI))- Def.
Human-level AI. The computer would need to be able to integrate many capabilites
Narrow AI-Def.
Ability to accomplish a narrow & well-defined set of goals (Matching resume to open job positions)
Cognition-Def.
the mental action or process of acquiring knowledge and understanding through thoughts, experience, and the senses. The psychological result of perception and learning and reasoning
Neural Networks -Def.
trained using machine learning approaches form the basis of most attempts at conginitc computing
Neuromorphic hardware -Def
It is being worked on in parallel to purely software.
The largest artificial neutral netwworks may have ca. 200 layern=> size of a frog brain
Artificial intelligence => ML=> DL
[AI] Human intelligence exhibited by machines - terms for (narrow) use cases, includes any simulation of human intelligence => [ML] Computer can learn from data without being explicitly programmes - terms for the technology, Uses numerical and statistical approaches (pattern), Models are built using “training” computation runs, can also train through usage => [DL aka Neural Network] A subfield of ML that uses specialized computational techniques, typically multi-layer (2+) artifical neural network, layering allows cascaded learning and abstraction levels (e.g. line recognition -> shape -> object -> scene ), Computationally intensive enabled by clouds, GPU’s, and increasingly more specialized HW
AI is becoming feasible because of a conflux of several technological advancements. The field is advancing rapidly beyond mathematical proofs because researchers can experiment with data and methods:
Powerful hardware -GPUS/TPUs accelerated training time of DNNs
Digitization & Big data- e.g. IOT, Digital receipts, Technologies like SAP HANA
Cloud computing -access powerful hardware with an Internet connection and credit card
Advances in applied deep neutral networks- Deep learning, new statistical techniques
What are First AI Attempts?
Rules-based “expert systems”:
If-then-else rules or algorithms, highly logical, structures & explainable, use structured, codified data
What is probabilistic machine learning
learn from data without being explicitly programmed (Cat-dog example)
Summary and Key Takeaways
- Intelligence is the ability to achieve complex goals. AI is intelligence exhibited by non-biological systems
- Currently the most feasible approach to AI is machine learning -> learning from large amount of data without beeing programmed
- AI leverages many different statistical modeling approaches from simple regressions to deep learning networks. Many are years old. Rules-based expert systems are still a useful and valid approach
- Feasible today because of ..(you know)
- People are comfy with rules-based systems but less comfortable with probabilistic
- we all train machine learning models every day. Big advantage
Basic ML Workflow
Trainings Data->Learning Algorithm -> Model1
Testing Data->Model2Accuracy Estimated
Model2->Model3 Prediction
Machine Learning is bases on 3 different categories of approaches to learning:
Supervised learning:
We provice the machine with examples of the desired solutiom (most common)
Unsupervised Learning: Data is set unlabeled- machine finds pattern /strucutes, Evaluation is qualitative or indirect
Reinforcement learning: Intelligence agents learn a polivy given neither data nor labels, only rules in their environment, observes environment, performs actions, gets rewards/penalties, Not to be confused with retraining a supervides learning model
Supervised learning:
In supervised the training inputs and desired outputs (called “labels”) are given by a “teacher”
Unsupervised learning
No labels are given to the learning algorithm, leaving it on its own to find the structure (costumer segments)
Reinforcement learning
we don’t start with data, but rather a defined envorinment. The agent trains its own policy throgh a process of trial and error
Classes of problem AI can adress
- Predicition, Forecasting, PAttern detection
- Computer Vision & Sensory Information
- Natural Language Processing and Unstructed Text
Regression (supervised) Def.
The output (i.e. what the machine predicts) is continours rather than discrete. Many kinds of regression models exist e.g. Year an employee will stay with the company
Classification (supervised) Def.
The model assigns inputs into 2 or more discrete classes. Inputs are aaigned to call with a given confidence level-Most machine learning adresses classificaiton problems e.g. employee will churn or not?, customer will choose product A oder Product B (or C)
Clustering Def. (mostly unsupervised) Def.
Divide the inputs into groups. The group are normally not know beforehand, making this typicaly an unsupervised task. E.g Customer Segmentation, Genome squencing
Dimensionality Reduction Def.
Map from higher dimensinal features space to lower dimensions is uses to combine several highly similar features into a singel feature E.g combine several items on a survey into a single “construct” variable
Examples for Models and Tools of statistical Model
Classification Analysis, Regression, Cluster analysis, Time series analysis, Association analysis, probability Distribution, Outlier Detection, Recommender Systems, Link prediction, Statistical Functions, Data prepation
How do you do feature engineering?
Creation of categorical or dummy variables, data cleaning, feature normaliazation /scaling, handle missing values, standardize input format
Misconception: A machine learns immediately after every click
Correction: A model is updated after we retrain model using the entire data set from scratch. Small exceptions include trransfer learning and active learning
Misconception: every action “reinfoces” and improves the model
Correction: reinforment learning is not training a supervised learning model. These are completely different approaches
Misconception: A model with the highest accuracy is the best model
Correction:Models with very high accuracy might be overfitted. When they are exposed to new data, te perfomance can degrade. Robustness and scability are equally imrportant.
Misconception: If we dont have training data, we can simulate data
If we know the functions and distribution already needed to simulate data, the we wouldnt need machine learning. Simulated data is viable in very rare circumstances (e.g. images created and roatetd using CAD software)
Misconception: Neural network are “superior” moddeling approaches
Neural Networks are appropriate (or the only possibility) for certain problem, but often “statistical” model are as accurate, more robust, and interpretable. Many models can be used to address the sam problem since performance depends on information in the data
Misconception: We require huge datasets to train a model
We don’t need until we test it. Some model train well with less than 100 others require more nach 100.000s. depends if the data contain a signal /on data quality
Misconception: If it doesn’t have 100% accuracy, we cannot use it
Nothing is 100% accurate” Unrealistiv expectation. AI resuts are probabilistic, exactly like results generated by humans (and AI is often fatser and more accurate..)
Misconception: We can explain all AI models
Many model types (esp, neural networks) are black boxes. Model interpretability is an active area of research in computer science - math problem rather than just engineering.
Misconception: We can generate actionable “insights” from the features in a machine learning model
AI interpretability (prediciton) is not explanatory modelling” Be careful.Two variables can be good predictors (correlation) but have the opposite or no relationsship in an explanatory model. Because X predicts Y is not to say if the change X it will Impact Y in the real world
Machine learning has several outcomes- which ones?
Regression, Classification, Clusterin and dimensionality reduction
What involves feature engineering
Cleaning, data, selecting relevant features, and codifying and tranforming the variables
Why are chatbots more feasible?
People are comfortable chatting with bots (messaging apps)
Performance of bots has increased substantially (largely thanks to neual networks /NLP)
Many benefits, Bots are: 24/7 aviable, efficient, cost reducing, multi-channel
What is the chief capability of conversational AI?
Assigning expressions (human sentences in unstructured text) to intens (categories of concepts that the bot understands)
Conversational AI - what are expressions ? What are intents?
Expressions: different sentences people say to communicate what they mean
Intents: Represent a concept that the bot will be able to understand by analyzing the user expressions, i.e. what the bot should know. Intents recognition is a core capability of a chatbot
What can NLP do?
NLP can recognize entities based on context and memory of the conversation. Simple to human but very difficult for computers. It extracts information from an expression (NER, sentiment analysis, language detection) and assigns it to an intent given confidence level. Responses are typically hard.-coded
