Ais machine learning
- 1. Artificial Immune Systems Andrew Watkins
- 2. Why the Immune System? • Recognition – Anomaly detection – Noise tolerance • Robustness • Feature extraction • Diversity • Reinforcement learning • Memory • Distributed • Multi-layered • Adaptive
- 3. Definition AIS are adaptive systems inspired by theoretical immunology and observed immune functions, principles and models, which are applied to complex problem domains (de Castro and Timmis)
- 4. Some History • Developed from the field of theoretical immunology in the mid 1980’s. – Suggested we ‘might look’ at the IS • 1990 – Bersini first use of immune algos to solve problems • Forrest et al – Computer Security mid 1990’s • Hunt et al, mid 1990’s – Machine learning
- 5. How does it work?
- 6. Immune Pattern Recognition • The immune recognition is based on the complementarity between the binding region of the receptor and a portion of the antigen called epitope. • Antibodies present a single type of receptor, antigens might present several epitopes. – This means that different antibodies can recognize a single antigen
- 7. Immune Responses Primary Response Secondary Response Cross-Reactive Response Antibody Concentration Lag Lag Response Response to Lag to Ag1 Ag1 + Ag3 ... Response to Ag1 Response to Ag2 ... ... ... Antigens Time Antigen Ag1 Antigen Ag1, Ag2 Ag1 + Ag3
- 9. Immune Network Theory • Idiotypic network (Jerne, 1974) • B cells co-stimulate each other – Treat each other a bit like antigens • Creates an immunological memory
- 10. Shape Space Formalism • Repertoire of the Vε × V ε immune system is Vε ε × × complete (Perelson, 1989) × × • Extensive regions of Vε ε × complementarity × • Some threshold of recognition
- 11. Self/Non-Self Recognition • Immune system needs to be able to differentiate between self and non-self cells • Antigenic encounters may result in cell death, therefore – Some kind of positive selection – Some element of negative selection
- 12. General Framework for AIS Solution Immune Algorithms Affinity Measures Representation Application Domain
- 13. Representation – Shape Space • Describe the general shape of a molecule •Describe interactions between molecules •Degree of binding between molecules •Complement threshold
- 14. Define their Interaction • Define the term Affinity • Affinity is related to distance L – Euclidian D= ∑ ( Abi − Ag i ) 2 i =1 • Other distance measures such as Hamming, Manhattan etc. etc. • Affinity Threshold
- 15. Basic Immune Models and Algorithms • Bone Marrow Models • Negative Selection Algorithms • Clonal Selection Algorithm • Somatic Hypermutation • Immune Network Models
- 16. Bone Marrow Models • Gene libraries are used to create antibodies from the bone marrow • Use this idea to generate attribute strings that represent receptors • Antibody production through a random concatenation from gene libraries
- 17. Negative Selection Algorithms • Forrest 1994: Idea taken from the negative selection of T-cells in the thymus • Applied initially to computer security • Split into two parts: – Censoring – Monitoring
- 18. Clonal Selection Algorithm (de Castro & von Zuben, 2001) Randomly initialise a population (P) For each pattern in Ag Determine affinity to each Ab in P Select n highest affinity from P Clone and mutate prop. to affinity with Ag Add new mutants to P endFor Select highest affinity Ab in P to form part of M Replace n number of random new ones Until stopping criteria
- 19. Immune Network Models (Timmis & Neal, 2001) Initialise the immune network (P) For each pattern in Ag Determine affinity to each Ab in P Calculate network interaction Allocate resources to the strongest members of P Remove weakest Ab in P EndFor If termination condition met exit else Clone and mutate each Ab in P (based on a given probability) Integrate new mutants into P based on affinity Repeat
- 20. Somatic Hypermutation • Mutation rate in proportion to affinity • Very controlled mutation in the natural immune system • The greater the antibody affinity the smaller its mutation rate • Classic trade-off between exploration and exploitation
- 21. How do AIS Compare? • Basic Components: – AIS B-cell in shape space (e.g. attribute strings) • Stimulation level – ANN Neuron • Activation function – GA chromosome • fitness
- 22. Comparing • Structure (Architecture) – AIS and GA fixed or variable sized populations, not connected in population based AIS – ANN and AIS • Do have network based AIS • ANN typically fixed structure (not always) • Learning takes place in weights in ANN
- 23. Comparing • Memory – AIS in B-cells • Network models in connections – ANN In weights of connections – GA individual chromosome
- 24. Comparing • Adaptation • Dynamics • Metadynamics • Interactions • Generalisation capabilities • Etc. many more.
- 25. Where are they used? • Dependable systems • Scheduling • Robotics • Security • Anomaly detection • Learning systems
- 26. Artificial Immune Recognition System (AIRS): An Immune-Inspired Supervised Learning Algorithm
- 27. AIRS: Immune Principles Employed • Clonal Selection • Based initially on immune networks, though found this did not work • Somatic hypermutation – Eventually • Recognition regions within shape space • Antibody/antigen binding
- 28. AIRS: Mapping from IS to AIS • Antibody Feature Vector • Recognition Combination of feature Ball (RB) vector and vector class • Antigens Training Data • Immune Memory Memory cells—set of mutated Artificial RBs
- 29. Classification • Stimulation of an ARB is based not only on its affinity to an antigen but also on its class when compared to the class of an antigen • Allocation of resources to the ARBs also takes into account the ARBs’ classifications when compared to the class of the antigen • Memory cell hyper-mutation and replacement is based primarily on classification and secondarily on affinity
- 30. AIRS Algorithm • Data normalization and initialization • Memory cell identification and ARB generation • Competition for resources in the development of a candidate memory cell • Potential introduction of the candidate memory cell into the set of established memory cells
- 31. Memory Cell Identification A Memory Cell Pool ARB Pool
- 32. MCmatch Found A 1 Memory Cell Pool MCmatch ARB Pool
- 33. ARB Generation A 1 Memory Cell Pool MCmatch Mutated Offspring 2 ARB Pool
- 34. Exposure of ARBs to Antigen A 1 Memory Cell Pool MCmatch Mutated Offspring 3 2 ARB Pool
- 35. Development of a Candidate Memory Cell A 1 Memory Cell Pool MCmatch Mutated Offspring 3 2 ARB Pool
- 36. Comparison of MCcandidate and MCmatch A 1 Memory Cell Pool MCmatch 4 A Mutated Offspring 3 2 MC candidate ARB Pool
- 37. Memory Cell Introduction A 1 Memory Cell Pool MCmatch 4 A 3 Mutated Offspring 5 2 MCcandidate ARB Pool
- 38. Memory Cells and Antigens
- 39. Memory Cells and Antigens
- 40. AIRS: Performance Evaluation Pima Indians Diabetes Fisher’s Iris Data Set Data Set Ionosphere Data Set Sonar Data Set
- 41. Iris Ionosphere Diabetes Sonar 1 Grobian 100% 3-NN + 98.7% Logdisc 77.7% TAP MFT 92.3% (rough) simplex Bayesian 2 SSV 98.0% 3-NN 96.7% IncNet 77.6% Naïve MFT Bayesian 90.4% 3 C-MLP2LN 98.0% IB3 96.7% DIPOL92 77.6% SVM 90.4% 4 PVM 2 rules 98.0% MLP + BP 96.0% Linear Discr. Anal. 77.5%- Best 2-layer MLP 90.4% 77.2% + BP, 12 hidden 5 PVM 1 rule 97.3% SMART 76.8% MLP+BP, 12 hidden 84.7% AIRS 94.9 6 C4.5 94.9% GTO DT (5xCV) 76.8% MLP+BP, 24 hidden 84.5% AIRS 96.7 7 FuNe-I 96.7% RIAC 94.6% ASI 76.6% 1-NN, Manhatten 84.2% 8 NEFCLASS 96.7% SVM 93.2% Fischer discr. anal 76.5% AIRS 84.0 9 CART 96.0% Non-linear 92.0% MLP+BP 76.4% MLP+BP, 6 83.5% perceptron hidden 10 FUNN 95.7% FSM + 92.8% LVQ 75.8% FSM - 83.6% rotation methodology? 11 1-NN 92.1% LFC 75.8% 1-NN Euclidean 82.2% 12 DB-CART 91.3% RBF 75.7% DB-CART, 10xCV 81.8% 13 Linear 90.7% NB 75.5- CART, 10xCV 67.9% perceptron 73.8% 14 OC1 DT 89.5% kNN, k=22, Manh 75.5% 15 CART 88.9% MML 75.5% … ... 22 AIRS 74.1 23 C4.5 73.0% 11 others reported with lower scores, including Bayes, Kohonen, kNN, ID3 …
- 42. AIRS: Observations • ARB Pool formulation was over complicated – Crude visualization – Memory only needs to be maintained in the Memory Cell Pool • Mutation Routine – Difference in Quality – Some redundancy
- 43. AIRS: Revisions • Memory Cell Evolution – Only Memory Cell Pool has different classes – ARB Pool only concerned with evolving memory cells • Somatic Hypermutation – Cell’s stimulation value indicates range of mutation possibilities – No longer need to mutate class
- 44. Comparisons: Classification Accuracy • Important to maintain accuracy AIRS1: Accuracy AIRS2: Accuracy Iris 96.7 96.0 Ionosphere 94.9 95.6 Diabetes 74.1 74.2 Sonar 84.0 84.9 • Why bother?
- 45. Comparisons: Data Reduction • Increase data reduction—increased efficiency Training Set Size AIRS1: Memory Cells AIRS2: Memory Cells Iris 120 42.1 / 65% 30.9 / 74% Ionosphere 200 140.7 / 30% 96.3 / 52% Diabetes 691 470.4 / 32% 273.4 / 60% Sonar 192 144.6 / 25% 177.7 / 7%
- 46. Features of AIRS • No need to know best architecture to get good results • Default settings within a few percent of the best it can get • User-adjustable parameters optimize performance for a given problem set • Generalization and data reduction
- 47. More Information • http://www.cs.ukc.ac.uk/people/rpg/abw5 • http://www.cs.ukc.ac.uk/people/staff/jt6 • http://www.cs.ukc.ac.uk/aisbook
Editor's Notes
- #3: Uniqueness : each individual possesses its own immune system, with its particular vulnerabilities and capabilities; Diversity : there is a large amount of types of elements (cells, molecules, proteins, etc.) that altogether perform the same role of protecting the body from malefic invaders. Additionally, there are different fronts of defense, like innate and adaptive immunity; Disposability ( robustness ): no single component of the natural immune system is essential for its functioning. Cell death is usually balanced by cell production; Autonomy : the immune system does not require outside management or maintenance. It autonomously classifies and eliminates pathogens, and it repairs itself by replacing damaged cells; Multilayered : multiple layers of different mechanisms are combined to provide high overall security, as summarized in Figure 2.5 (Section 2.3); No secure layer : any cell of the human body can be attacked by the immune system, including those of the immune system itself; Recognition of foreigners : the (harmful) molecules that are not native to the body are recognized and eliminated by the immune system; Anomaly detection : the immune system can detect and react to pathogens that the body has never encountered before; Dynamically changing coverage : as the immune system can not maintain a set of cells and molecules large enough to detect all pathogens, it makes a trade-off between space and time. It maintains a circulating pool of lymphocytes that is constantly being changed through cell death, production and reproduction; Distributability : the immune cells, molecules and organs are distributed all over the body and, most importantly, are not subject to any centralized control; Imperfect detection ( noise tolerance ): an absolute recognition of the pathogens is not required, hence the system is flexible; Reinforcement learning and memory : the immune system can “learn” the structures of pathogens. It retains the ability to recognize previously seen pathogens through immune memory, so that future responses to the same pathogens are faster and stronger; and An arms race : the vertebrate immune system replicates cells to deal with replicating pathogens, otherwise the pathogens would quickly overwhelm the immune defenses.
- #10: Mention Bersinis' principles
- #19: 1. Randomly initialize a population of individuals ( P ); 2. For each pattern of S , present it to the population P and determine its affinity with each element of the population P ; 3. Select n 1 highest affinity elements of P and generate copies of these individuals proportionally to their affinity with the antigen. The higher the affinity, the higher the number of copies, and vice-versa; 4. Mutate all these copies with a rate proportional to their affinity with the input pattern: the higher the affinity, the smaller the mutation rate, and vice-versa; 5. Add these mutated individuals to the population P and re-select n 2 of these maturated (optimized) individuals to be kept as the memory M of the system; 6. Replace a number n 3 of individuals with low affinity by (randomly generated) new ones; 7. Repeat Steps 2 to 6 until a certain stopping criterion is met.
- #20: Initialise the immune network (P) For each pattern in Ag Determine affinity to each P’ Calculate network interaction Allocate resources to the strongest members of P Remove weakest P EndFor If termination condition met exit else Clone and mutate each P (based on probability a) Integrate new mutants into P based on affinity Repeat
- #30: Sparse in AIS literature Not as straight forward as initially suspected
- #33: MCmatch is found
- #34: New ARBs are generated to be put into the population
- #35: The competition for system wide resources The use of mutation for diversification and shape-space exploration The use of an average stimulation threshold as a criterion for determining when to stop training on a given antigen
- #36: The competition for system wide resources The use of mutation for diversification and shape-space exploration The use of an average stimulation threshold as a criterion for determining when to stop training on a given antigen
- #37: Compare response of MCmatch and MCcandidate to the antigen. Compare the affinity value of MCmatch and MCcandidate to each other
- #38: Introduce the just-developed candidate memory cell, MCcandidate , into the set of existing memory cells MC Replace MCmatch The evolved memory cells are available for use for classification.
- #41: Iris: 3 way classification problem; 150 data items; 5XCV; avg. 3 times; 4 features Ionosphere: 2-way classification, good & bad radar returns; 34 features ; 200 in training, 151 test set Diabetes: 2-way class, has diabetes or not; 10XCV; 8 features ; 768 instances total Sonar: 2-way class; 13XCV; 60 features ; 16 instances in each test set